gpkg.slim.models 0.5.1

TF-Slim models (Guild AI)

gpkg.slim.models
################

*TF-Slim models (Guild AI)*

Models
######

images
======

*Generic images dataset*

Operations
^^^^^^^^^^

prepare
-------

*Prepare images for training*

Flags
`````

**images**
*Directory containing images to prepare (required)*

**random-seed**
*Seed used for train/validation split (randomly generated)*

**val-split**
*Percentage of images reserved for validation (30)*


inception
=========

*TF-Slim Inception v1 classifier*

Operations
^^^^^^^^^^

evaluate
--------

*Evaluate a trained model*

Flags
`````

**batch-size**
*Number of examples in each evaluated batch (100)*

**eval-batches**
*Number of batches to evaluate (all available)*

**step**
*Checkpoint step to evaluate (latest checkpoint)*

export-and-freeze
-----------------

*Export an inference graph with checkpoint weights*

Flags
`````

**step**
*Checkpoint step to use for the frozen graph (latest checkpoint)*

finetune
--------

*Finetune a trained model*

Flags
`````

**auto-scale**
*Adjust applicable flags for multi-GPU systems (yes)

Set to 'no' to disable any flag value adjustments.

When this value is 'yes' (the default) the following flags are adjusted on
multi-GPU systems:

- clones

- learning-rate

`clones` is set to the number of available GPUs.

`learning-rate` is adjusted by multiplying its specified value by the
number of GPUs.

Flags are not adjusted on single GPU or CPU only systems.

*

**batch-size**
*Number of examples in each training batch (32)*

**clones**
*Number of model clones (calculated)

This value is automatically set to the number of available GPUs if `auto-
scale` is 'yes'.

When `auto-scale` is 'no' this value can be increased from 1 to train the
model in parallel on multiple GPUs.

*

**learning-rate**
*Initial learning rate (0.0001)*

**learning-rate-decay-epochs**
*Number of epochs after which learning rate decays (2.0)*

**learning-rate-decay-factor**
*Learning rate decay factor (0.94)*

**learning-rate-decay-type**
*Method used to decay the learning rate (exponential)

Choices:
exponential
fixed
polynomial

*

**learning-rate-end**
*Minimal learning rate used by polynomial learning rate decay (0.0001)*

**log-save-seconds**
*Frequency of log summary saves in seconds (60)*

**log-steps**
*Frequency of summary logs in steps (100)*

**model-save-seconds**
*Frequency of model saves (checkpoints) in seconds (600)*

**optimizer**
*Optimizer used to train (rmsprop)

Choices:
adadelta
adagrad
adam
ftrl
momentum
rmsprop
sgd

*

**preprocessing**
*Preprocessing to use (default for model)*

**preprocessors**
*Number of preprocessing threads (calculated)

This value is automatically set to logical CPU count / 2 if `auto-scale`
is 'yes'.

When `auto-scale` is 'no' this value can be set to optimize the
preprocessor thread count for the system.

*

**readers**
*Number of parallel data readers (calculated)

This value is automatically set to logical CPU count / 2 if `auto-scale`
is 'yes'.

When `auto-scale` is 'no' this value can be set to optimize data reader
performance for the system.

*

**train-steps**
*Number of steps to train (train indefinitely)*

**weight-decay**
*Decay on the model weights (4e-05)*

label
-----

*Classify an image using a trained model*

Flags
`````

**image**
*Path to image to classify (required)*

tflite
------

*Generate a TFLite file from a frozen graph*

Flags
`````

**output-format**
*TF Lite output format (tflite)

Choices:
tflite
graphviz_dot

*

**quantized**
*Whether or not output arrays are quantized (no)

Choices:
yes
no

*

**quantized-inputs**
*Whether or not input arrays are quantized (no)

Choices:
yes
no

*

train
-----

*Train model from scratch*

Flags
`````

**auto-scale**
*Adjust applicable flags for multi-GPU systems (yes)

Set to 'no' to disable any flag value adjustments.

When this value is 'yes' (the default) the following flags are adjusted on
multi-GPU systems:

- clones

- learning-rate

`clones` is set to the number of available GPUs.

`learning-rate` is adjusted by multiplying its specified value by the
number of GPUs.

Flags are not adjusted on single GPU or CPU only systems.

*

**batch-size**
*Number of examples in each training batch (32)*

**clones**
*Number of model clones (calculated)

This value is automatically set to the number of available GPUs if `auto-
scale` is 'yes'.

When `auto-scale` is 'no' this value can be increased from 1 to train the
model in parallel on multiple GPUs.

*

**learning-rate**
*Initial learning rate (0.001)*

**learning-rate-decay-epochs**
*Number of epochs after which learning rate decays (2.0)*

**learning-rate-decay-factor**
*Learning rate decay factor (0.94)*

**learning-rate-decay-type**
*Method used to decay the learning rate (exponential)

Choices:
exponential
fixed
polynomial

*

**learning-rate-end**
*Minimal learning rate used by polynomial learning rate decay (0.0001)*

**log-save-seconds**
*Frequency of log summary saves in seconds (60)*

**log-steps**
*Frequency of summary logs in steps (100)*

**model-save-seconds**
*Frequency of model saves (checkpoints) in seconds (600)*

**optimizer**
*Optimizer used to train (rmsprop)

Choices:
adadelta
adagrad
adam
ftrl
momentum
rmsprop
sgd

*

**preprocessing**
*Preprocessing to use (default for model)*

**preprocessors**
*Number of preprocessing threads (calculated)

This value is automatically set to logical CPU count / 2 if `auto-scale`
is 'yes'.

When `auto-scale` is 'no' this value can be set to optimize the
preprocessor thread count for the system.

*

**readers**
*Number of parallel data readers (calculated)

This value is automatically set to logical CPU count / 2 if `auto-scale`
is 'yes'.

When `auto-scale` is 'no' this value can be set to optimize data reader
performance for the system.

*

**train-steps**
*Number of steps to train (train indefinitely)*

**weight-decay**
*Decay on the model weights (4e-05)*

transfer-learn
--------------

*Train model using transfer learning*

Flags
`````

**auto-scale**
*Adjust applicable flags for multi-GPU systems (yes)

Set to 'no' to disable any flag value adjustments.

When this value is 'yes' (the default) the following flags are adjusted on
multi-GPU systems:

- clones

- learning-rate

`clones` is set to the number of available GPUs.

`learning-rate` is adjusted by multiplying its specified value by the
number of GPUs.

Flags are not adjusted on single GPU or CPU only systems.

*

**batch-size**
*Number of examples in each training batch (32)*

**clones**
*Number of model clones (calculated)

This value is automatically set to the number of available GPUs if `auto-
scale` is 'yes'.

When `auto-scale` is 'no' this value can be increased from 1 to train the
model in parallel on multiple GPUs.

*

**learning-rate**
*Initial learning rate (0.001)*

**learning-rate-decay-epochs**
*Number of epochs after which learning rate decays (2.0)*

**learning-rate-decay-factor**
*Learning rate decay factor (0.94)*

**learning-rate-decay-type**
*Method used to decay the learning rate (exponential)

Choices:
exponential
fixed
polynomial

*

**learning-rate-end**
*Minimal learning rate used by polynomial learning rate decay (0.0001)*

**log-save-seconds**
*Frequency of log summary saves in seconds (60)*

**log-steps**
*Frequency of summary logs in steps (100)*

**model-save-seconds**
*Frequency of model saves (checkpoints) in seconds (600)*

**optimizer**
*Optimizer used to train (rmsprop)

Choices:
adadelta
adagrad
adam
ftrl
momentum
rmsprop
sgd

*

**preprocessing**
*Preprocessing to use (default for model)*

**preprocessors**
*Number of preprocessing threads (calculated)

This value is automatically set to logical CPU count / 2 if `auto-scale`
is 'yes'.

When `auto-scale` is 'no' this value can be set to optimize the
preprocessor thread count for the system.

*

**readers**
*Number of parallel data readers (calculated)

This value is automatically set to logical CPU count / 2 if `auto-scale`
is 'yes'.

When `auto-scale` is 'no' this value can be set to optimize data reader
performance for the system.

*

**train-steps**
*Number of steps to train (train indefinitely)*

**weight-decay**
*Decay on the model weights (4e-05)*


inception-resnet-v2
===================

*TF-Slim Inception ResNet v2 classifier*

Operations
^^^^^^^^^^

evaluate
--------

*Evaluate a trained model*

Flags
`````

**batch-size**
*Number of examples in each evaluated batch (100)*

**eval-batches**
*Number of batches to evaluate (all available)*

**step**
*Checkpoint step to evaluate (latest checkpoint)*

export-and-freeze
-----------------

*Export an inference graph with checkpoint weights*

Flags
`````

**step**
*Checkpoint step to use for the frozen graph (latest checkpoint)*

finetune
--------

*Finetune a trained model*

Flags
`````

**auto-scale**
*Adjust applicable flags for multi-GPU systems (yes)

Set to 'no' to disable any flag value adjustments.

When this value is 'yes' (the default) the following flags are adjusted on
multi-GPU systems:

- clones

- learning-rate

`clones` is set to the number of available GPUs.

`learning-rate` is adjusted by multiplying its specified value by the
number of GPUs.

Flags are not adjusted on single GPU or CPU only systems.

*

**batch-size**
*Number of examples in each training batch (32)*

**clones**
*Number of model clones (calculated)

This value is automatically set to the number of available GPUs if `auto-
scale` is 'yes'.

When `auto-scale` is 'no' this value can be increased from 1 to train the
model in parallel on multiple GPUs.

*

**learning-rate**
*Initial learning rate (0.0001)*

**learning-rate-decay-epochs**
*Number of epochs after which learning rate decays (2.0)*

**learning-rate-decay-factor**
*Learning rate decay factor (0.94)*

**learning-rate-decay-type**
*Method used to decay the learning rate (exponential)

Choices:
exponential
fixed
polynomial

*

**learning-rate-end**
*Minimal learning rate used by polynomial learning rate decay (0.0001)*

**log-save-seconds**
*Frequency of log summary saves in seconds (60)*

**log-steps**
*Frequency of summary logs in steps (100)*

**model-save-seconds**
*Frequency of model saves (checkpoints) in seconds (600)*

**optimizer**
*Optimizer used to train (rmsprop)

Choices:
adadelta
adagrad
adam
ftrl
momentum
rmsprop
sgd

*

**preprocessing**
*Preprocessing to use (default for model)*

**preprocessors**
*Number of preprocessing threads (calculated)

This value is automatically set to logical CPU count / 2 if `auto-scale`
is 'yes'.

When `auto-scale` is 'no' this value can be set to optimize the
preprocessor thread count for the system.

*

**readers**
*Number of parallel data readers (calculated)

This value is automatically set to logical CPU count / 2 if `auto-scale`
is 'yes'.

When `auto-scale` is 'no' this value can be set to optimize data reader
performance for the system.

*

**train-steps**
*Number of steps to train (train indefinitely)*

**weight-decay**
*Decay on the model weights (4e-05)*

label
-----

*Classify an image using a trained model*

Flags
`````

**image**
*Path to image to classify (required)*

tflite
------

*Generate a TFLite file from a frozen graph*

Flags
`````

**output-format**
*TF Lite output format (tflite)

Choices:
tflite
graphviz_dot

*

**quantized**
*Whether or not output arrays are quantized (no)

Choices:
yes
no

*

**quantized-inputs**
*Whether or not input arrays are quantized (no)

Choices:
yes
no

*

train
-----

*Train model from scratch*

Flags
`````

**auto-scale**
*Adjust applicable flags for multi-GPU systems (yes)

Set to 'no' to disable any flag value adjustments.

When this value is 'yes' (the default) the following flags are adjusted on
multi-GPU systems:

- clones

- learning-rate

`clones` is set to the number of available GPUs.

`learning-rate` is adjusted by multiplying its specified value by the
number of GPUs.

Flags are not adjusted on single GPU or CPU only systems.

*

**batch-size**
*Number of examples in each training batch (32)*

**clones**
*Number of model clones (calculated)

This value is automatically set to the number of available GPUs if `auto-
scale` is 'yes'.

When `auto-scale` is 'no' this value can be increased from 1 to train the
model in parallel on multiple GPUs.

*

**learning-rate**
*Initial learning rate (0.001)*

**learning-rate-decay-epochs**
*Number of epochs after which learning rate decays (2.0)*

**learning-rate-decay-factor**
*Learning rate decay factor (0.94)*

**learning-rate-decay-type**
*Method used to decay the learning rate (exponential)

Choices:
exponential
fixed
polynomial

*

**learning-rate-end**
*Minimal learning rate used by polynomial learning rate decay (0.0001)*

**log-save-seconds**
*Frequency of log summary saves in seconds (60)*

**log-steps**
*Frequency of summary logs in steps (100)*

**model-save-seconds**
*Frequency of model saves (checkpoints) in seconds (600)*

**optimizer**
*Optimizer used to train (rmsprop)

Choices:
adadelta
adagrad
adam
ftrl
momentum
rmsprop
sgd

*

**preprocessing**
*Preprocessing to use (default for model)*

**preprocessors**
*Number of preprocessing threads (calculated)

This value is automatically set to logical CPU count / 2 if `auto-scale`
is 'yes'.

When `auto-scale` is 'no' this value can be set to optimize the
preprocessor thread count for the system.

*

**readers**
*Number of parallel data readers (calculated)

This value is automatically set to logical CPU count / 2 if `auto-scale`
is 'yes'.

When `auto-scale` is 'no' this value can be set to optimize data reader
performance for the system.

*

**train-steps**
*Number of steps to train (train indefinitely)*

**weight-decay**
*Decay on the model weights (4e-05)*

transfer-learn
--------------

*Train model using transfer learning*

Flags
`````

**auto-scale**
*Adjust applicable flags for multi-GPU systems (yes)

Set to 'no' to disable any flag value adjustments.

When this value is 'yes' (the default) the following flags are adjusted on
multi-GPU systems:

- clones

- learning-rate

`clones` is set to the number of available GPUs.

`learning-rate` is adjusted by multiplying its specified value by the
number of GPUs.

Flags are not adjusted on single GPU or CPU only systems.

*

**batch-size**
*Number of examples in each training batch (32)*

**clones**
*Number of model clones (calculated)

This value is automatically set to the number of available GPUs if `auto-
scale` is 'yes'.

When `auto-scale` is 'no' this value can be increased from 1 to train the
model in parallel on multiple GPUs.

*

**learning-rate**
*Initial learning rate (0.001)*

**learning-rate-decay-epochs**
*Number of epochs after which learning rate decays (2.0)*

**learning-rate-decay-factor**
*Learning rate decay factor (0.94)*

**learning-rate-decay-type**
*Method used to decay the learning rate (exponential)

Choices:
exponential
fixed
polynomial

*

**learning-rate-end**
*Minimal learning rate used by polynomial learning rate decay (0.0001)*

**log-save-seconds**
*Frequency of log summary saves in seconds (60)*

**log-steps**
*Frequency of summary logs in steps (100)*

**model-save-seconds**
*Frequency of model saves (checkpoints) in seconds (600)*

**optimizer**
*Optimizer used to train (rmsprop)

Choices:
adadelta
adagrad
adam
ftrl
momentum
rmsprop
sgd

*

**preprocessing**
*Preprocessing to use (default for model)*

**preprocessors**
*Number of preprocessing threads (calculated)

This value is automatically set to logical CPU count / 2 if `auto-scale`
is 'yes'.

When `auto-scale` is 'no' this value can be set to optimize the
preprocessor thread count for the system.

*

**readers**
*Number of parallel data readers (calculated)

This value is automatically set to logical CPU count / 2 if `auto-scale`
is 'yes'.

When `auto-scale` is 'no' this value can be set to optimize data reader
performance for the system.

*

**train-steps**
*Number of steps to train (train indefinitely)*

**weight-decay**
*Decay on the model weights (4e-05)*


inception-v2
============

*TF-Slim Inception v2 classifier*

Operations
^^^^^^^^^^

evaluate
--------

*Evaluate a trained model*

Flags
`````

**batch-size**
*Number of examples in each evaluated batch (100)*

**eval-batches**
*Number of batches to evaluate (all available)*

**step**
*Checkpoint step to evaluate (latest checkpoint)*

export-and-freeze
-----------------

*Export an inference graph with checkpoint weights*

Flags
`````

**step**
*Checkpoint step to use for the frozen graph (latest checkpoint)*

finetune
--------

*Finetune a trained model*

Flags
`````

**auto-scale**
*Adjust applicable flags for multi-GPU systems (yes)

Set to 'no' to disable any flag value adjustments.

When this value is 'yes' (the default) the following flags are adjusted on
multi-GPU systems:

- clones

- learning-rate

`clones` is set to the number of available GPUs.

`learning-rate` is adjusted by multiplying its specified value by the
number of GPUs.

Flags are not adjusted on single GPU or CPU only systems.

*

**batch-size**
*Number of examples in each training batch (32)*

**clones**
*Number of model clones (calculated)

This value is automatically set to the number of available GPUs if `auto-
scale` is 'yes'.

When `auto-scale` is 'no' this value can be increased from 1 to train the
model in parallel on multiple GPUs.

*

**learning-rate**
*Initial learning rate (0.0001)*

**learning-rate-decay-epochs**
*Number of epochs after which learning rate decays (2.0)*

**learning-rate-decay-factor**
*Learning rate decay factor (0.94)*

**learning-rate-decay-type**
*Method used to decay the learning rate (exponential)

Choices:
exponential
fixed
polynomial

*

**learning-rate-end**
*Minimal learning rate used by polynomial learning rate decay (0.0001)*

**log-save-seconds**
*Frequency of log summary saves in seconds (60)*

**log-steps**
*Frequency of summary logs in steps (100)*

**model-save-seconds**
*Frequency of model saves (checkpoints) in seconds (600)*

**optimizer**
*Optimizer used to train (rmsprop)

Choices:
adadelta
adagrad
adam
ftrl
momentum
rmsprop
sgd

*

**preprocessing**
*Preprocessing to use (default for model)*

**preprocessors**
*Number of preprocessing threads (calculated)

This value is automatically set to logical CPU count / 2 if `auto-scale`
is 'yes'.

When `auto-scale` is 'no' this value can be set to optimize the
preprocessor thread count for the system.

*

**readers**
*Number of parallel data readers (calculated)

This value is automatically set to logical CPU count / 2 if `auto-scale`
is 'yes'.

When `auto-scale` is 'no' this value can be set to optimize data reader
performance for the system.

*

**train-steps**
*Number of steps to train (train indefinitely)*

**weight-decay**
*Decay on the model weights (4e-05)*

label
-----

*Classify an image using a trained model*

Flags
`````

**image**
*Path to image to classify (required)*

tflite
------

*Generate a TFLite file from a frozen graph*

Flags
`````

**output-format**
*TF Lite output format (tflite)

Choices:
tflite
graphviz_dot

*

**quantized**
*Whether or not output arrays are quantized (no)

Choices:
yes
no

*

**quantized-inputs**
*Whether or not input arrays are quantized (no)

Choices:
yes
no

*

train
-----

*Train model from scratch*

Flags
`````

**auto-scale**
*Adjust applicable flags for multi-GPU systems (yes)

Set to 'no' to disable any flag value adjustments.

When this value is 'yes' (the default) the following flags are adjusted on
multi-GPU systems:

- clones

- learning-rate

`clones` is set to the number of available GPUs.

`learning-rate` is adjusted by multiplying its specified value by the
number of GPUs.

Flags are not adjusted on single GPU or CPU only systems.

*

**batch-size**
*Number of examples in each training batch (32)*

**clones**
*Number of model clones (calculated)

This value is automatically set to the number of available GPUs if `auto-
scale` is 'yes'.

When `auto-scale` is 'no' this value can be increased from 1 to train the
model in parallel on multiple GPUs.

*

**learning-rate**
*Initial learning rate (0.001)*

**learning-rate-decay-epochs**
*Number of epochs after which learning rate decays (2.0)*

**learning-rate-decay-factor**
*Learning rate decay factor (0.94)*

**learning-rate-decay-type**
*Method used to decay the learning rate (exponential)

Choices:
exponential
fixed
polynomial

*

**learning-rate-end**
*Minimal learning rate used by polynomial learning rate decay (0.0001)*

**log-save-seconds**
*Frequency of log summary saves in seconds (60)*

**log-steps**
*Frequency of summary logs in steps (100)*

**model-save-seconds**
*Frequency of model saves (checkpoints) in seconds (600)*

**optimizer**
*Optimizer used to train (rmsprop)

Choices:
adadelta
adagrad
adam
ftrl
momentum
rmsprop
sgd

*

**preprocessing**
*Preprocessing to use (default for model)*

**preprocessors**
*Number of preprocessing threads (calculated)

This value is automatically set to logical CPU count / 2 if `auto-scale`
is 'yes'.

When `auto-scale` is 'no' this value can be set to optimize the
preprocessor thread count for the system.

*

**readers**
*Number of parallel data readers (calculated)

This value is automatically set to logical CPU count / 2 if `auto-scale`
is 'yes'.

When `auto-scale` is 'no' this value can be set to optimize data reader
performance for the system.

*

**train-steps**
*Number of steps to train (train indefinitely)*

**weight-decay**
*Decay on the model weights (4e-05)*

transfer-learn
--------------

*Train model using transfer learning*

Flags
`````

**auto-scale**
*Adjust applicable flags for multi-GPU systems (yes)

Set to 'no' to disable any flag value adjustments.

When this value is 'yes' (the default) the following flags are adjusted on
multi-GPU systems:

- clones

- learning-rate

`clones` is set to the number of available GPUs.

`learning-rate` is adjusted by multiplying its specified value by the
number of GPUs.

Flags are not adjusted on single GPU or CPU only systems.

*

**batch-size**
*Number of examples in each training batch (32)*

**clones**
*Number of model clones (calculated)

This value is automatically set to the number of available GPUs if `auto-
scale` is 'yes'.

When `auto-scale` is 'no' this value can be increased from 1 to train the
model in parallel on multiple GPUs.

*

**learning-rate**
*Initial learning rate (0.001)*

**learning-rate-decay-epochs**
*Number of epochs after which learning rate decays (2.0)*

**learning-rate-decay-factor**
*Learning rate decay factor (0.94)*

**learning-rate-decay-type**
*Method used to decay the learning rate (exponential)

Choices:
exponential
fixed
polynomial

*

**learning-rate-end**
*Minimal learning rate used by polynomial learning rate decay (0.0001)*

**log-save-seconds**
*Frequency of log summary saves in seconds (60)*

**log-steps**
*Frequency of summary logs in steps (100)*

**model-save-seconds**
*Frequency of model saves (checkpoints) in seconds (600)*

**optimizer**
*Optimizer used to train (rmsprop)

Choices:
adadelta
adagrad
adam
ftrl
momentum
rmsprop
sgd

*

**preprocessing**
*Preprocessing to use (default for model)*

**preprocessors**
*Number of preprocessing threads (calculated)

This value is automatically set to logical CPU count / 2 if `auto-scale`
is 'yes'.

When `auto-scale` is 'no' this value can be set to optimize the
preprocessor thread count for the system.

*

**readers**
*Number of parallel data readers (calculated)

This value is automatically set to logical CPU count / 2 if `auto-scale`
is 'yes'.

When `auto-scale` is 'no' this value can be set to optimize data reader
performance for the system.

*

**train-steps**
*Number of steps to train (train indefinitely)*

**weight-decay**
*Decay on the model weights (4e-05)*


inception-v3
============

*TF-Slim Inception v3 classifier*

Operations
^^^^^^^^^^

evaluate
--------

*Evaluate a trained model*

Flags
`````

**batch-size**
*Number of examples in each evaluated batch (100)*

**eval-batches**
*Number of batches to evaluate (all available)*

**step**
*Checkpoint step to evaluate (latest checkpoint)*

export-and-freeze
-----------------

*Export an inference graph with checkpoint weights*

Flags
`````

**step**
*Checkpoint step to use for the frozen graph (latest checkpoint)*

finetune
--------

*Finetune a trained model*

Flags
`````

**auto-scale**
*Adjust applicable flags for multi-GPU systems (yes)

Set to 'no' to disable any flag value adjustments.

When this value is 'yes' (the default) the following flags are adjusted on
multi-GPU systems:

- clones

- learning-rate

`clones` is set to the number of available GPUs.

`learning-rate` is adjusted by multiplying its specified value by the
number of GPUs.

Flags are not adjusted on single GPU or CPU only systems.

*

**batch-size**
*Number of examples in each training batch (32)*

**clones**
*Number of model clones (calculated)

This value is automatically set to the number of available GPUs if `auto-
scale` is 'yes'.

When `auto-scale` is 'no' this value can be increased from 1 to train the
model in parallel on multiple GPUs.

*

**learning-rate**
*Initial learning rate (0.0001)*

**learning-rate-decay-epochs**
*Number of epochs after which learning rate decays (2.0)*

**learning-rate-decay-factor**
*Learning rate decay factor (0.94)*

**learning-rate-decay-type**
*Method used to decay the learning rate (exponential)

Choices:
exponential
fixed
polynomial

*

**learning-rate-end**
*Minimal learning rate used by polynomial learning rate decay (0.0001)*

**log-save-seconds**
*Frequency of log summary saves in seconds (60)*

**log-steps**
*Frequency of summary logs in steps (100)*

**model-save-seconds**
*Frequency of model saves (checkpoints) in seconds (600)*

**optimizer**
*Optimizer used to train (rmsprop)

Choices:
adadelta
adagrad
adam
ftrl
momentum
rmsprop
sgd

*

**preprocessing**
*Preprocessing to use (default for model)*

**preprocessors**
*Number of preprocessing threads (calculated)

This value is automatically set to logical CPU count / 2 if `auto-scale`
is 'yes'.

When `auto-scale` is 'no' this value can be set to optimize the
preprocessor thread count for the system.

*

**readers**
*Number of parallel data readers (calculated)

This value is automatically set to logical CPU count / 2 if `auto-scale`
is 'yes'.

When `auto-scale` is 'no' this value can be set to optimize data reader
performance for the system.

*

**train-steps**
*Number of steps to train (train indefinitely)*

**weight-decay**
*Decay on the model weights (4e-05)*

label
-----

*Classify an image using a trained model*

Flags
`````

**image**
*Path to image to classify (required)*

tflite
------

*Generate a TFLite file from a frozen graph*

Flags
`````

**output-format**
*TF Lite output format (tflite)

Choices:
tflite
graphviz_dot

*

**quantized**
*Whether or not output arrays are quantized (no)

Choices:
yes
no

*

**quantized-inputs**
*Whether or not input arrays are quantized (no)

Choices:
yes
no

*

train
-----

*Train model from scratch*

Flags
`````

**auto-scale**
*Adjust applicable flags for multi-GPU systems (yes)

Set to 'no' to disable any flag value adjustments.

When this value is 'yes' (the default) the following flags are adjusted on
multi-GPU systems:

- clones

- learning-rate

`clones` is set to the number of available GPUs.

`learning-rate` is adjusted by multiplying its specified value by the
number of GPUs.

Flags are not adjusted on single GPU or CPU only systems.

*

**batch-size**
*Number of examples in each training batch (32)*

**clones**
*Number of model clones (calculated)

This value is automatically set to the number of available GPUs if `auto-
scale` is 'yes'.

When `auto-scale` is 'no' this value can be increased from 1 to train the
model in parallel on multiple GPUs.

*

**learning-rate**
*Initial learning rate (0.001)*

**learning-rate-decay-epochs**
*Number of epochs after which learning rate decays (2.0)*

**learning-rate-decay-factor**
*Learning rate decay factor (0.94)*

**learning-rate-decay-type**
*Method used to decay the learning rate (exponential)

Choices:
exponential
fixed
polynomial

*

**learning-rate-end**
*Minimal learning rate used by polynomial learning rate decay (0.0001)*

**log-save-seconds**
*Frequency of log summary saves in seconds (60)*

**log-steps**
*Frequency of summary logs in steps (100)*

**model-save-seconds**
*Frequency of model saves (checkpoints) in seconds (600)*

**optimizer**
*Optimizer used to train (rmsprop)

Choices:
adadelta
adagrad
adam
ftrl
momentum
rmsprop
sgd

*

**preprocessing**
*Preprocessing to use (default for model)*

**preprocessors**
*Number of preprocessing threads (calculated)

This value is automatically set to logical CPU count / 2 if `auto-scale`
is 'yes'.

When `auto-scale` is 'no' this value can be set to optimize the
preprocessor thread count for the system.

*

**readers**
*Number of parallel data readers (calculated)

This value is automatically set to logical CPU count / 2 if `auto-scale`
is 'yes'.

When `auto-scale` is 'no' this value can be set to optimize data reader
performance for the system.

*

**train-steps**
*Number of steps to train (train indefinitely)*

**weight-decay**
*Decay on the model weights (4e-05)*

transfer-learn
--------------

*Train model using transfer learning*

Flags
`````

**auto-scale**
*Adjust applicable flags for multi-GPU systems (yes)

Set to 'no' to disable any flag value adjustments.

When this value is 'yes' (the default) the following flags are adjusted on
multi-GPU systems:

- clones

- learning-rate

`clones` is set to the number of available GPUs.

`learning-rate` is adjusted by multiplying its specified value by the
number of GPUs.

Flags are not adjusted on single GPU or CPU only systems.

*

**batch-size**
*Number of examples in each training batch (32)*

**clones**
*Number of model clones (calculated)

This value is automatically set to the number of available GPUs if `auto-
scale` is 'yes'.

When `auto-scale` is 'no' this value can be increased from 1 to train the
model in parallel on multiple GPUs.

*

**learning-rate**
*Initial learning rate (0.001)*

**learning-rate-decay-epochs**
*Number of epochs after which learning rate decays (2.0)*

**learning-rate-decay-factor**
*Learning rate decay factor (0.94)*

**learning-rate-decay-type**
*Method used to decay the learning rate (exponential)

Choices:
exponential
fixed
polynomial

*

**learning-rate-end**
*Minimal learning rate used by polynomial learning rate decay (0.0001)*

**log-save-seconds**
*Frequency of log summary saves in seconds (60)*

**log-steps**
*Frequency of summary logs in steps (100)*

**model-save-seconds**
*Frequency of model saves (checkpoints) in seconds (600)*

**optimizer**
*Optimizer used to train (rmsprop)

Choices:
adadelta
adagrad
adam
ftrl
momentum
rmsprop
sgd

*

**preprocessing**
*Preprocessing to use (default for model)*

**preprocessors**
*Number of preprocessing threads (calculated)

This value is automatically set to logical CPU count / 2 if `auto-scale`
is 'yes'.

When `auto-scale` is 'no' this value can be set to optimize the
preprocessor thread count for the system.

*

**readers**
*Number of parallel data readers (calculated)

This value is automatically set to logical CPU count / 2 if `auto-scale`
is 'yes'.

When `auto-scale` is 'no' this value can be set to optimize data reader
performance for the system.

*

**train-steps**
*Number of steps to train (train indefinitely)*

**weight-decay**
*Decay on the model weights (4e-05)*


inception-v4
============

*TF-Slim Inception v4 classifier*

Operations
^^^^^^^^^^

evaluate
--------

*Evaluate a trained model*

Flags
`````

**batch-size**
*Number of examples in each evaluated batch (100)*

**eval-batches**
*Number of batches to evaluate (all available)*

**step**
*Checkpoint step to evaluate (latest checkpoint)*

export-and-freeze
-----------------

*Export an inference graph with checkpoint weights*

Flags
`````

**step**
*Checkpoint step to use for the frozen graph (latest checkpoint)*

finetune
--------

*Finetune a trained model*

Flags
`````

**auto-scale**
*Adjust applicable flags for multi-GPU systems (yes)

Set to 'no' to disable any flag value adjustments.

When this value is 'yes' (the default) the following flags are adjusted on
multi-GPU systems:

- clones

- learning-rate

`clones` is set to the number of available GPUs.

`learning-rate` is adjusted by multiplying its specified value by the
number of GPUs.

Flags are not adjusted on single GPU or CPU only systems.

*

**batch-size**
*Number of examples in each training batch (32)*

**clones**
*Number of model clones (calculated)

This value is automatically set to the number of available GPUs if `auto-
scale` is 'yes'.

When `auto-scale` is 'no' this value can be increased from 1 to train the
model in parallel on multiple GPUs.

*

**learning-rate**
*Initial learning rate (0.0001)*

**learning-rate-decay-epochs**
*Number of epochs after which learning rate decays (2.0)*

**learning-rate-decay-factor**
*Learning rate decay factor (0.94)*

**learning-rate-decay-type**
*Method used to decay the learning rate (exponential)

Choices:
exponential
fixed
polynomial

*

**learning-rate-end**
*Minimal learning rate used by polynomial learning rate decay (0.0001)*

**log-save-seconds**
*Frequency of log summary saves in seconds (60)*

**log-steps**
*Frequency of summary logs in steps (100)*

**model-save-seconds**
*Frequency of model saves (checkpoints) in seconds (600)*

**optimizer**
*Optimizer used to train (rmsprop)

Choices:
adadelta
adagrad
adam
ftrl
momentum
rmsprop
sgd

*

**preprocessing**
*Preprocessing to use (default for model)*

**preprocessors**
*Number of preprocessing threads (calculated)

This value is automatically set to logical CPU count / 2 if `auto-scale`
is 'yes'.

When `auto-scale` is 'no' this value can be set to optimize the
preprocessor thread count for the system.

*

**readers**
*Number of parallel data readers (calculated)

This value is automatically set to logical CPU count / 2 if `auto-scale`
is 'yes'.

When `auto-scale` is 'no' this value can be set to optimize data reader
performance for the system.

*

**train-steps**
*Number of steps to train (train indefinitely)*

**weight-decay**
*Decay on the model weights (4e-05)*

label
-----

*Classify an image using a trained model*

Flags
`````

**image**
*Path to image to classify (required)*

tflite
------

*Generate a TFLite file from a frozen graph*

Flags
`````

**output-format**
*TF Lite output format (tflite)

Choices:
tflite
graphviz_dot

*

**quantized**
*Whether or not output arrays are quantized (no)

Choices:
yes
no

*

**quantized-inputs**
*Whether or not input arrays are quantized (no)

Choices:
yes
no

*

train
-----

*Train model from scratch*

Flags
`````

**auto-scale**
*Adjust applicable flags for multi-GPU systems (yes)

Set to 'no' to disable any flag value adjustments.

When this value is 'yes' (the default) the following flags are adjusted on
multi-GPU systems:

- clones

- learning-rate

`clones` is set to the number of available GPUs.

`learning-rate` is adjusted by multiplying its specified value by the
number of GPUs.

Flags are not adjusted on single GPU or CPU only systems.

*

**batch-size**
*Number of examples in each training batch (32)*

**clones**
*Number of model clones (calculated)

This value is automatically set to the number of available GPUs if `auto-
scale` is 'yes'.

When `auto-scale` is 'no' this value can be increased from 1 to train the
model in parallel on multiple GPUs.

*

**learning-rate**
*Initial learning rate (0.001)*

**learning-rate-decay-epochs**
*Number of epochs after which learning rate decays (2.0)*

**learning-rate-decay-factor**
*Learning rate decay factor (0.94)*

**learning-rate-decay-type**
*Method used to decay the learning rate (exponential)

Choices:
exponential
fixed
polynomial

*

**learning-rate-end**
*Minimal learning rate used by polynomial learning rate decay (0.0001)*

**log-save-seconds**
*Frequency of log summary saves in seconds (60)*

**log-steps**
*Frequency of summary logs in steps (100)*

**model-save-seconds**
*Frequency of model saves (checkpoints) in seconds (600)*

**optimizer**
*Optimizer used to train (rmsprop)

Choices:
adadelta
adagrad
adam
ftrl
momentum
rmsprop
sgd

*

**preprocessing**
*Preprocessing to use (default for model)*

**preprocessors**
*Number of preprocessing threads (calculated)

This value is automatically set to logical CPU count / 2 if `auto-scale`
is 'yes'.

When `auto-scale` is 'no' this value can be set to optimize the
preprocessor thread count for the system.

*

**readers**
*Number of parallel data readers (calculated)

This value is automatically set to logical CPU count / 2 if `auto-scale`
is 'yes'.

When `auto-scale` is 'no' this value can be set to optimize data reader
performance for the system.

*

**train-steps**
*Number of steps to train (train indefinitely)*

**weight-decay**
*Decay on the model weights (4e-05)*

transfer-learn
--------------

*Train model using transfer learning*

Flags
`````

**auto-scale**
*Adjust applicable flags for multi-GPU systems (yes)

Set to 'no' to disable any flag value adjustments.

When this value is 'yes' (the default) the following flags are adjusted on
multi-GPU systems:

- clones

- learning-rate

`clones` is set to the number of available GPUs.

`learning-rate` is adjusted by multiplying its specified value by the
number of GPUs.

Flags are not adjusted on single GPU or CPU only systems.

*

**batch-size**
*Number of examples in each training batch (32)*

**clones**
*Number of model clones (calculated)

This value is automatically set to the number of available GPUs if `auto-
scale` is 'yes'.

When `auto-scale` is 'no' this value can be increased from 1 to train the
model in parallel on multiple GPUs.

*

**learning-rate**
*Initial learning rate (0.001)*

**learning-rate-decay-epochs**
*Number of epochs after which learning rate decays (2.0)*

**learning-rate-decay-factor**
*Learning rate decay factor (0.94)*

**learning-rate-decay-type**
*Method used to decay the learning rate (exponential)

Choices:
exponential
fixed
polynomial

*

**learning-rate-end**
*Minimal learning rate used by polynomial learning rate decay (0.0001)*

**log-save-seconds**
*Frequency of log summary saves in seconds (60)*

**log-steps**
*Frequency of summary logs in steps (100)*

**model-save-seconds**
*Frequency of model saves (checkpoints) in seconds (600)*

**optimizer**
*Optimizer used to train (rmsprop)

Choices:
adadelta
adagrad
adam
ftrl
momentum
rmsprop
sgd

*

**preprocessing**
*Preprocessing to use (default for model)*

**preprocessors**
*Number of preprocessing threads (calculated)

This value is automatically set to logical CPU count / 2 if `auto-scale`
is 'yes'.

When `auto-scale` is 'no' this value can be set to optimize the
preprocessor thread count for the system.

*

**readers**
*Number of parallel data readers (calculated)

This value is automatically set to logical CPU count / 2 if `auto-scale`
is 'yes'.

When `auto-scale` is 'no' this value can be set to optimize data reader
performance for the system.

*

**train-steps**
*Number of steps to train (train indefinitely)*

**weight-decay**
*Decay on the model weights (4e-05)*


mobilenet
=========

*TF-Slim Mobilenet v1 classifier*

Operations
^^^^^^^^^^

evaluate
--------

*Evaluate a trained model*

Flags
`````

**batch-size**
*Number of examples in each evaluated batch (100)*

**eval-batches**
*Number of batches to evaluate (all available)*

**step**
*Checkpoint step to evaluate (latest checkpoint)*

export-and-freeze
-----------------

*Export an inference graph with checkpoint weights*

Flags
`````

**step**
*Checkpoint step to use for the frozen graph (latest checkpoint)*

finetune
--------

*Finetune a trained model*

Flags
`````

**auto-scale**
*Adjust applicable flags for multi-GPU systems (yes)

Set to 'no' to disable any flag value adjustments.

When this value is 'yes' (the default) the following flags are adjusted on
multi-GPU systems:

- clones

- learning-rate

`clones` is set to the number of available GPUs.

`learning-rate` is adjusted by multiplying its specified value by the
number of GPUs.

Flags are not adjusted on single GPU or CPU only systems.

*

**batch-size**
*Number of examples in each training batch (32)*

**clones**
*Number of model clones (calculated)

This value is automatically set to the number of available GPUs if `auto-
scale` is 'yes'.

When `auto-scale` is 'no' this value can be increased from 1 to train the
model in parallel on multiple GPUs.

*

**learning-rate**
*Initial learning rate (0.0001)*

**learning-rate-decay-epochs**
*Number of epochs after which learning rate decays (2.0)*

**learning-rate-decay-factor**
*Learning rate decay factor (0.94)*

**learning-rate-decay-type**
*Method used to decay the learning rate (exponential)

Choices:
exponential
fixed
polynomial

*

**learning-rate-end**
*Minimal learning rate used by polynomial learning rate decay (0.0001)*

**log-save-seconds**
*Frequency of log summary saves in seconds (60)*

**log-steps**
*Frequency of summary logs in steps (100)*

**model-save-seconds**
*Frequency of model saves (checkpoints) in seconds (600)*

**optimizer**
*Optimizer used to train (rmsprop)

Choices:
adadelta
adagrad
adam
ftrl
momentum
rmsprop
sgd

*

**preprocessing**
*Preprocessing to use (default for model)*

**preprocessors**
*Number of preprocessing threads (calculated)

This value is automatically set to logical CPU count / 2 if `auto-scale`
is 'yes'.

When `auto-scale` is 'no' this value can be set to optimize the
preprocessor thread count for the system.

*

**readers**
*Number of parallel data readers (calculated)

This value is automatically set to logical CPU count / 2 if `auto-scale`
is 'yes'.

When `auto-scale` is 'no' this value can be set to optimize data reader
performance for the system.

*

**train-steps**
*Number of steps to train (train indefinitely)*

**weight-decay**
*Decay on the model weights (4e-05)*

label
-----

*Classify an image using a trained model*

Flags
`````

**image**
*Path to image to classify (required)*

tflite
------

*Generate a TFLite file from a frozen graph*

Flags
`````

**output-format**
*TF Lite output format (tflite)

Choices:
tflite
graphviz_dot

*

**quantized**
*Whether or not output arrays are quantized (no)

Choices:
yes
no

*

**quantized-inputs**
*Whether or not input arrays are quantized (no)

Choices:
yes
no

*

train
-----

*Train model from scratch*

Flags
`````

**auto-scale**
*Adjust applicable flags for multi-GPU systems (yes)

Set to 'no' to disable any flag value adjustments.

When this value is 'yes' (the default) the following flags are adjusted on
multi-GPU systems:

- clones

- learning-rate

`clones` is set to the number of available GPUs.

`learning-rate` is adjusted by multiplying its specified value by the
number of GPUs.

Flags are not adjusted on single GPU or CPU only systems.

*

**batch-size**
*Number of examples in each training batch (32)*

**clones**
*Number of model clones (calculated)

This value is automatically set to the number of available GPUs if `auto-
scale` is 'yes'.

When `auto-scale` is 'no' this value can be increased from 1 to train the
model in parallel on multiple GPUs.

*

**learning-rate**
*Initial learning rate (0.001)*

**learning-rate-decay-epochs**
*Number of epochs after which learning rate decays (2.0)*

**learning-rate-decay-factor**
*Learning rate decay factor (0.94)*

**learning-rate-decay-type**
*Method used to decay the learning rate (exponential)

Choices:
exponential
fixed
polynomial

*

**learning-rate-end**
*Minimal learning rate used by polynomial learning rate decay (0.0001)*

**log-save-seconds**
*Frequency of log summary saves in seconds (60)*

**log-steps**
*Frequency of summary logs in steps (100)*

**model-save-seconds**
*Frequency of model saves (checkpoints) in seconds (600)*

**optimizer**
*Optimizer used to train (rmsprop)

Choices:
adadelta
adagrad
adam
ftrl
momentum
rmsprop
sgd

*

**preprocessing**
*Preprocessing to use (default for model)*

**preprocessors**
*Number of preprocessing threads (calculated)

This value is automatically set to logical CPU count / 2 if `auto-scale`
is 'yes'.

When `auto-scale` is 'no' this value can be set to optimize the
preprocessor thread count for the system.

*

**readers**
*Number of parallel data readers (calculated)

This value is automatically set to logical CPU count / 2 if `auto-scale`
is 'yes'.

When `auto-scale` is 'no' this value can be set to optimize data reader
performance for the system.

*

**train-steps**
*Number of steps to train (train indefinitely)*

**weight-decay**
*Decay on the model weights (4e-05)*

transfer-learn
--------------

*Train model using transfer learning*

Flags
`````

**auto-scale**
*Adjust applicable flags for multi-GPU systems (yes)

Set to 'no' to disable any flag value adjustments.

When this value is 'yes' (the default) the following flags are adjusted on
multi-GPU systems:

- clones

- learning-rate

`clones` is set to the number of available GPUs.

`learning-rate` is adjusted by multiplying its specified value by the
number of GPUs.

Flags are not adjusted on single GPU or CPU only systems.

*

**batch-size**
*Number of examples in each training batch (32)*

**clones**
*Number of model clones (calculated)

This value is automatically set to the number of available GPUs if `auto-
scale` is 'yes'.

When `auto-scale` is 'no' this value can be increased from 1 to train the
model in parallel on multiple GPUs.

*

**learning-rate**
*Initial learning rate (0.001)*

**learning-rate-decay-epochs**
*Number of epochs after which learning rate decays (2.0)*

**learning-rate-decay-factor**
*Learning rate decay factor (0.94)*

**learning-rate-decay-type**
*Method used to decay the learning rate (exponential)

Choices:
exponential
fixed
polynomial

*

**learning-rate-end**
*Minimal learning rate used by polynomial learning rate decay (0.0001)*

**log-save-seconds**
*Frequency of log summary saves in seconds (60)*

**log-steps**
*Frequency of summary logs in steps (100)*

**model-save-seconds**
*Frequency of model saves (checkpoints) in seconds (600)*

**optimizer**
*Optimizer used to train (rmsprop)

Choices:
adadelta
adagrad
adam
ftrl
momentum
rmsprop
sgd

*

**preprocessing**
*Preprocessing to use (default for model)*

**preprocessors**
*Number of preprocessing threads (calculated)

This value is automatically set to logical CPU count / 2 if `auto-scale`
is 'yes'.

When `auto-scale` is 'no' this value can be set to optimize the
preprocessor thread count for the system.

*

**readers**
*Number of parallel data readers (calculated)

This value is automatically set to logical CPU count / 2 if `auto-scale`
is 'yes'.

When `auto-scale` is 'no' this value can be set to optimize data reader
performance for the system.

*

**train-steps**
*Number of steps to train (train indefinitely)*

**weight-decay**
*Decay on the model weights (4e-05)*


mobilenet-v2-1.4
================

*TF-Slim Mobilenet v2 classifier*

Operations
^^^^^^^^^^

evaluate
--------

*Evaluate a trained model*

Flags
`````

**batch-size**
*Number of examples in each evaluated batch (100)*

**eval-batches**
*Number of batches to evaluate (all available)*

**step**
*Checkpoint step to evaluate (latest checkpoint)*

export-and-freeze
-----------------

*Export an inference graph with checkpoint weights*

Flags
`````

**step**
*Checkpoint step to use for the frozen graph (latest checkpoint)*

finetune
--------

*Finetune a trained model*

Flags
`````

**auto-scale**
*Adjust applicable flags for multi-GPU systems (yes)

Set to 'no' to disable any flag value adjustments.

When this value is 'yes' (the default) the following flags are adjusted on
multi-GPU systems:

- clones

- learning-rate

`clones` is set to the number of available GPUs.

`learning-rate` is adjusted by multiplying its specified value by the
number of GPUs.

Flags are not adjusted on single GPU or CPU only systems.

*

**batch-size**
*Number of examples in each training batch (32)*

**clones**
*Number of model clones (calculated)

This value is automatically set to the number of available GPUs if `auto-
scale` is 'yes'.

When `auto-scale` is 'no' this value can be increased from 1 to train the
model in parallel on multiple GPUs.

*

**learning-rate**
*Initial learning rate (0.0001)*

**learning-rate-decay-epochs**
*Number of epochs after which learning rate decays (2.0)*

**learning-rate-decay-factor**
*Learning rate decay factor (0.94)*

**learning-rate-decay-type**
*Method used to decay the learning rate (exponential)

Choices:
exponential
fixed
polynomial

*

**learning-rate-end**
*Minimal learning rate used by polynomial learning rate decay (0.0001)*

**log-save-seconds**
*Frequency of log summary saves in seconds (60)*

**log-steps**
*Frequency of summary logs in steps (100)*

**model-save-seconds**
*Frequency of model saves (checkpoints) in seconds (600)*

**optimizer**
*Optimizer used to train (rmsprop)

Choices:
adadelta
adagrad
adam
ftrl
momentum
rmsprop
sgd

*

**preprocessing**
*Preprocessing to use (default for model)*

**preprocessors**
*Number of preprocessing threads (calculated)

This value is automatically set to logical CPU count / 2 if `auto-scale`
is 'yes'.

When `auto-scale` is 'no' this value can be set to optimize the
preprocessor thread count for the system.

*

**readers**
*Number of parallel data readers (calculated)

This value is automatically set to logical CPU count / 2 if `auto-scale`
is 'yes'.

When `auto-scale` is 'no' this value can be set to optimize data reader
performance for the system.

*

**train-steps**
*Number of steps to train (train indefinitely)*

**weight-decay**
*Decay on the model weights (4e-05)*

label
-----

*Classify an image using a trained model*

Flags
`````

**image**
*Path to image to classify (required)*

tflite
------

*Generate a TFLite file from a frozen graph*

Flags
`````

**output-format**
*TF Lite output format (tflite)

Choices:
tflite
graphviz_dot

*

**quantized**
*Whether or not output arrays are quantized (no)

Choices:
yes
no

*

**quantized-inputs**
*Whether or not input arrays are quantized (no)

Choices:
yes
no

*

train
-----

*Train model from scratch*

Flags
`````

**auto-scale**
*Adjust applicable flags for multi-GPU systems (yes)

Set to 'no' to disable any flag value adjustments.

When this value is 'yes' (the default) the following flags are adjusted on
multi-GPU systems:

- clones

- learning-rate

`clones` is set to the number of available GPUs.

`learning-rate` is adjusted by multiplying its specified value by the
number of GPUs.

Flags are not adjusted on single GPU or CPU only systems.

*

**batch-size**
*Number of examples in each training batch (32)*

**clones**
*Number of model clones (calculated)

This value is automatically set to the number of available GPUs if `auto-
scale` is 'yes'.

When `auto-scale` is 'no' this value can be increased from 1 to train the
model in parallel on multiple GPUs.

*

**learning-rate**
*Initial learning rate (0.001)*

**learning-rate-decay-epochs**
*Number of epochs after which learning rate decays (2.0)*

**learning-rate-decay-factor**
*Learning rate decay factor (0.94)*

**learning-rate-decay-type**
*Method used to decay the learning rate (exponential)

Choices:
exponential
fixed
polynomial

*

**learning-rate-end**
*Minimal learning rate used by polynomial learning rate decay (0.0001)*

**log-save-seconds**
*Frequency of log summary saves in seconds (60)*

**log-steps**
*Frequency of summary logs in steps (100)*

**model-save-seconds**
*Frequency of model saves (checkpoints) in seconds (600)*

**optimizer**
*Optimizer used to train (rmsprop)

Choices:
adadelta
adagrad
adam
ftrl
momentum
rmsprop
sgd

*

**preprocessing**
*Preprocessing to use (default for model)*

**preprocessors**
*Number of preprocessing threads (calculated)

This value is automatically set to logical CPU count / 2 if `auto-scale`
is 'yes'.

When `auto-scale` is 'no' this value can be set to optimize the
preprocessor thread count for the system.

*

**readers**
*Number of parallel data readers (calculated)

This value is automatically set to logical CPU count / 2 if `auto-scale`
is 'yes'.

When `auto-scale` is 'no' this value can be set to optimize data reader
performance for the system.

*

**train-steps**
*Number of steps to train (train indefinitely)*

**weight-decay**
*Decay on the model weights (4e-05)*

transfer-learn
--------------

*Train model using transfer learning*

Flags
`````

**auto-scale**
*Adjust applicable flags for multi-GPU systems (yes)

Set to 'no' to disable any flag value adjustments.

When this value is 'yes' (the default) the following flags are adjusted on
multi-GPU systems:

- clones

- learning-rate

`clones` is set to the number of available GPUs.

`learning-rate` is adjusted by multiplying its specified value by the
number of GPUs.

Flags are not adjusted on single GPU or CPU only systems.

*

**batch-size**
*Number of examples in each training batch (32)*

**clones**
*Number of model clones (calculated)

This value is automatically set to the number of available GPUs if `auto-
scale` is 'yes'.

When `auto-scale` is 'no' this value can be increased from 1 to train the
model in parallel on multiple GPUs.

*

**learning-rate**
*Initial learning rate (0.001)*

**learning-rate-decay-epochs**
*Number of epochs after which learning rate decays (2.0)*

**learning-rate-decay-factor**
*Learning rate decay factor (0.94)*

**learning-rate-decay-type**
*Method used to decay the learning rate (exponential)

Choices:
exponential
fixed
polynomial

*

**learning-rate-end**
*Minimal learning rate used by polynomial learning rate decay (0.0001)*

**log-save-seconds**
*Frequency of log summary saves in seconds (60)*

**log-steps**
*Frequency of summary logs in steps (100)*

**model-save-seconds**
*Frequency of model saves (checkpoints) in seconds (600)*

**optimizer**
*Optimizer used to train (rmsprop)

Choices:
adadelta
adagrad
adam
ftrl
momentum
rmsprop
sgd

*

**preprocessing**
*Preprocessing to use (default for model)*

**preprocessors**
*Number of preprocessing threads (calculated)

This value is automatically set to logical CPU count / 2 if `auto-scale`
is 'yes'.

When `auto-scale` is 'no' this value can be set to optimize the
preprocessor thread count for the system.

*

**readers**
*Number of parallel data readers (calculated)

This value is automatically set to logical CPU count / 2 if `auto-scale`
is 'yes'.

When `auto-scale` is 'no' this value can be set to optimize data reader
performance for the system.

*

**train-steps**
*Number of steps to train (train indefinitely)*

**weight-decay**
*Decay on the model weights (4e-05)*


nasnet-large
============

*TF-Slim NASNet large classifier*

Operations
^^^^^^^^^^

evaluate
--------

*Evaluate a trained model*

Flags
`````

**batch-size**
*Number of examples in each evaluated batch (100)*

**eval-batches**
*Number of batches to evaluate (all available)*

**step**
*Checkpoint step to evaluate (latest checkpoint)*

export-and-freeze
-----------------

*Export an inference graph with checkpoint weights*

Flags
`````

**step**
*Checkpoint step to use for the frozen graph (latest checkpoint)*

finetune
--------

*Finetune a trained model*

Flags
`````

**auto-scale**
*Adjust applicable flags for multi-GPU systems (yes)

Set to 'no' to disable any flag value adjustments.

When this value is 'yes' (the default) the following flags are adjusted on
multi-GPU systems:

- clones

- learning-rate

`clones` is set to the number of available GPUs.

`learning-rate` is adjusted by multiplying its specified value by the
number of GPUs.

Flags are not adjusted on single GPU or CPU only systems.

*

**batch-size**
*Number of examples in each training batch (32)*

**clones**
*Number of model clones (calculated)

This value is automatically set to the number of available GPUs if `auto-
scale` is 'yes'.

When `auto-scale` is 'no' this value can be increased from 1 to train the
model in parallel on multiple GPUs.

*

**learning-rate**
*Initial learning rate (0.0001)*

**learning-rate-decay-epochs**
*Number of epochs after which learning rate decays (2.0)*

**learning-rate-decay-factor**
*Learning rate decay factor (0.94)*

**learning-rate-decay-type**
*Method used to decay the learning rate (exponential)

Choices:
exponential
fixed
polynomial

*

**learning-rate-end**
*Minimal learning rate used by polynomial learning rate decay (0.0001)*

**log-save-seconds**
*Frequency of log summary saves in seconds (60)*

**log-steps**
*Frequency of summary logs in steps (100)*

**model-save-seconds**
*Frequency of model saves (checkpoints) in seconds (600)*

**optimizer**
*Optimizer used to train (rmsprop)

Choices:
adadelta
adagrad
adam
ftrl
momentum
rmsprop
sgd

*

**preprocessing**
*Preprocessing to use (default for model)*

**preprocessors**
*Number of preprocessing threads (calculated)

This value is automatically set to logical CPU count / 2 if `auto-scale`
is 'yes'.

When `auto-scale` is 'no' this value can be set to optimize the
preprocessor thread count for the system.

*

**readers**
*Number of parallel data readers (calculated)

This value is automatically set to logical CPU count / 2 if `auto-scale`
is 'yes'.

When `auto-scale` is 'no' this value can be set to optimize data reader
performance for the system.

*

**train-steps**
*Number of steps to train (train indefinitely)*

**weight-decay**
*Decay on the model weights (4e-05)*

label
-----

*Classify an image using a trained model*

Flags
`````

**image**
*Path to image to classify (required)*

tflite
------

*Generate a TFLite file from a frozen graph*

Flags
`````

**output-format**
*TF Lite output format (tflite)

Choices:
tflite
graphviz_dot

*

**quantized**
*Whether or not output arrays are quantized (no)

Choices:
yes
no

*

**quantized-inputs**
*Whether or not input arrays are quantized (no)

Choices:
yes
no

*

train
-----

*Train model from scratch*

Flags
`````

**auto-scale**
*Adjust applicable flags for multi-GPU systems (yes)

Set to 'no' to disable any flag value adjustments.

When this value is 'yes' (the default) the following flags are adjusted on
multi-GPU systems:

- clones

- learning-rate

`clones` is set to the number of available GPUs.

`learning-rate` is adjusted by multiplying its specified value by the
number of GPUs.

Flags are not adjusted on single GPU or CPU only systems.

*

**batch-size**
*Number of examples in each training batch (32)*

**clones**
*Number of model clones (calculated)

This value is automatically set to the number of available GPUs if `auto-
scale` is 'yes'.

When `auto-scale` is 'no' this value can be increased from 1 to train the
model in parallel on multiple GPUs.

*

**learning-rate**
*Initial learning rate (0.001)*

**learning-rate-decay-epochs**
*Number of epochs after which learning rate decays (2.0)*

**learning-rate-decay-factor**
*Learning rate decay factor (0.94)*

**learning-rate-decay-type**
*Method used to decay the learning rate (exponential)

Choices:
exponential
fixed
polynomial

*

**learning-rate-end**
*Minimal learning rate used by polynomial learning rate decay (0.0001)*

**log-save-seconds**
*Frequency of log summary saves in seconds (60)*

**log-steps**
*Frequency of summary logs in steps (100)*

**model-save-seconds**
*Frequency of model saves (checkpoints) in seconds (600)*

**optimizer**
*Optimizer used to train (rmsprop)

Choices:
adadelta
adagrad
adam
ftrl
momentum
rmsprop
sgd

*

**preprocessing**
*Preprocessing to use (default for model)*

**preprocessors**
*Number of preprocessing threads (calculated)

This value is automatically set to logical CPU count / 2 if `auto-scale`
is 'yes'.

When `auto-scale` is 'no' this value can be set to optimize the
preprocessor thread count for the system.

*

**readers**
*Number of parallel data readers (calculated)

This value is automatically set to logical CPU count / 2 if `auto-scale`
is 'yes'.

When `auto-scale` is 'no' this value can be set to optimize data reader
performance for the system.

*

**train-steps**
*Number of steps to train (train indefinitely)*

**weight-decay**
*Decay on the model weights (4e-05)*

transfer-learn
--------------

*Train model using transfer learning*

Flags
`````

**auto-scale**
*Adjust applicable flags for multi-GPU systems (yes)

Set to 'no' to disable any flag value adjustments.

When this value is 'yes' (the default) the following flags are adjusted on
multi-GPU systems:

- clones

- learning-rate

`clones` is set to the number of available GPUs.

`learning-rate` is adjusted by multiplying its specified value by the
number of GPUs.

Flags are not adjusted on single GPU or CPU only systems.

*

**batch-size**
*Number of examples in each training batch (32)*

**clones**
*Number of model clones (calculated)

This value is automatically set to the number of available GPUs if `auto-
scale` is 'yes'.

When `auto-scale` is 'no' this value can be increased from 1 to train the
model in parallel on multiple GPUs.

*

**learning-rate**
*Initial learning rate (0.001)*

**learning-rate-decay-epochs**
*Number of epochs after which learning rate decays (2.0)*

**learning-rate-decay-factor**
*Learning rate decay factor (0.94)*

**learning-rate-decay-type**
*Method used to decay the learning rate (exponential)

Choices:
exponential
fixed
polynomial

*

**learning-rate-end**
*Minimal learning rate used by polynomial learning rate decay (0.0001)*

**log-save-seconds**
*Frequency of log summary saves in seconds (60)*

**log-steps**
*Frequency of summary logs in steps (100)*

**model-save-seconds**
*Frequency of model saves (checkpoints) in seconds (600)*

**optimizer**
*Optimizer used to train (rmsprop)

Choices:
adadelta
adagrad
adam
ftrl
momentum
rmsprop
sgd

*

**preprocessing**
*Preprocessing to use (default for model)*

**preprocessors**
*Number of preprocessing threads (calculated)

This value is automatically set to logical CPU count / 2 if `auto-scale`
is 'yes'.

When `auto-scale` is 'no' this value can be set to optimize the
preprocessor thread count for the system.

*

**readers**
*Number of parallel data readers (calculated)

This value is automatically set to logical CPU count / 2 if `auto-scale`
is 'yes'.

When `auto-scale` is 'no' this value can be set to optimize data reader
performance for the system.

*

**train-steps**
*Number of steps to train (train indefinitely)*

**weight-decay**
*Decay on the model weights (4e-05)*


nasnet-mobile
=============

*TF-Slim NASNet mobile classifier*

Operations
^^^^^^^^^^

evaluate
--------

*Evaluate a trained model*

Flags
`````

**batch-size**
*Number of examples in each evaluated batch (100)*

**eval-batches**
*Number of batches to evaluate (all available)*

**step**
*Checkpoint step to evaluate (latest checkpoint)*

export-and-freeze
-----------------

*Export an inference graph with checkpoint weights*

Flags
`````

**step**
*Checkpoint step to use for the frozen graph (latest checkpoint)*

finetune
--------

*Finetune a trained model*

Flags
`````

**auto-scale**
*Adjust applicable flags for multi-GPU systems (yes)

Set to 'no' to disable any flag value adjustments.

When this value is 'yes' (the default) the following flags are adjusted on
multi-GPU systems:

- clones

- learning-rate

`clones` is set to the number of available GPUs.

`learning-rate` is adjusted by multiplying its specified value by the
number of GPUs.

Flags are not adjusted on single GPU or CPU only systems.

*

**batch-size**
*Number of examples in each training batch (32)*

**clones**
*Number of model clones (calculated)

This value is automatically set to the number of available GPUs if `auto-
scale` is 'yes'.

When `auto-scale` is 'no' this value can be increased from 1 to train the
model in parallel on multiple GPUs.

*

**learning-rate**
*Initial learning rate (0.0001)*

**learning-rate-decay-epochs**
*Number of epochs after which learning rate decays (2.0)*

**learning-rate-decay-factor**
*Learning rate decay factor (0.94)*

**learning-rate-decay-type**
*Method used to decay the learning rate (exponential)

Choices:
exponential
fixed
polynomial

*

**learning-rate-end**
*Minimal learning rate used by polynomial learning rate decay (0.0001)*

**log-save-seconds**
*Frequency of log summary saves in seconds (60)*

**log-steps**
*Frequency of summary logs in steps (100)*

**model-save-seconds**
*Frequency of model saves (checkpoints) in seconds (600)*

**optimizer**
*Optimizer used to train (rmsprop)

Choices:
adadelta
adagrad
adam
ftrl
momentum
rmsprop
sgd

*

**preprocessing**
*Preprocessing to use (default for model)*

**preprocessors**
*Number of preprocessing threads (calculated)

This value is automatically set to logical CPU count / 2 if `auto-scale`
is 'yes'.

When `auto-scale` is 'no' this value can be set to optimize the
preprocessor thread count for the system.

*

**readers**
*Number of parallel data readers (calculated)

This value is automatically set to logical CPU count / 2 if `auto-scale`
is 'yes'.

When `auto-scale` is 'no' this value can be set to optimize data reader
performance for the system.

*

**train-steps**
*Number of steps to train (train indefinitely)*

**weight-decay**
*Decay on the model weights (4e-05)*

label
-----

*Classify an image using a trained model*

Flags
`````

**image**
*Path to image to classify (required)*

tflite
------

*Generate a TFLite file from a frozen graph*

Flags
`````

**output-format**
*TF Lite output format (tflite)

Choices:
tflite
graphviz_dot

*

**quantized**
*Whether or not output arrays are quantized (no)

Choices:
yes
no

*

**quantized-inputs**
*Whether or not input arrays are quantized (no)

Choices:
yes
no

*

train
-----

*Train model from scratch*

Flags
`````

**auto-scale**
*Adjust applicable flags for multi-GPU systems (yes)

Set to 'no' to disable any flag value adjustments.

When this value is 'yes' (the default) the following flags are adjusted on
multi-GPU systems:

- clones

- learning-rate

`clones` is set to the number of available GPUs.

`learning-rate` is adjusted by multiplying its specified value by the
number of GPUs.

Flags are not adjusted on single GPU or CPU only systems.

*

**batch-size**
*Number of examples in each training batch (32)*

**clones**
*Number of model clones (calculated)

This value is automatically set to the number of available GPUs if `auto-
scale` is 'yes'.

When `auto-scale` is 'no' this value can be increased from 1 to train the
model in parallel on multiple GPUs.

*

**learning-rate**
*Initial learning rate (0.001)*

**learning-rate-decay-epochs**
*Number of epochs after which learning rate decays (2.0)*

**learning-rate-decay-factor**
*Learning rate decay factor (0.94)*

**learning-rate-decay-type**
*Method used to decay the learning rate (exponential)

Choices:
exponential
fixed
polynomial

*

**learning-rate-end**
*Minimal learning rate used by polynomial learning rate decay (0.0001)*

**log-save-seconds**
*Frequency of log summary saves in seconds (60)*

**log-steps**
*Frequency of summary logs in steps (100)*

**model-save-seconds**
*Frequency of model saves (checkpoints) in seconds (600)*

**optimizer**
*Optimizer used to train (rmsprop)

Choices:
adadelta
adagrad
adam
ftrl
momentum
rmsprop
sgd

*

**preprocessing**
*Preprocessing to use (default for model)*

**preprocessors**
*Number of preprocessing threads (calculated)

This value is automatically set to logical CPU count / 2 if `auto-scale`
is 'yes'.

When `auto-scale` is 'no' this value can be set to optimize the
preprocessor thread count for the system.

*

**readers**
*Number of parallel data readers (calculated)

This value is automatically set to logical CPU count / 2 if `auto-scale`
is 'yes'.

When `auto-scale` is 'no' this value can be set to optimize data reader
performance for the system.

*

**train-steps**
*Number of steps to train (train indefinitely)*

**weight-decay**
*Decay on the model weights (4e-05)*

transfer-learn
--------------

*Train model using transfer learning*

Flags
`````

**auto-scale**
*Adjust applicable flags for multi-GPU systems (yes)

Set to 'no' to disable any flag value adjustments.

When this value is 'yes' (the default) the following flags are adjusted on
multi-GPU systems:

- clones

- learning-rate

`clones` is set to the number of available GPUs.

`learning-rate` is adjusted by multiplying its specified value by the
number of GPUs.

Flags are not adjusted on single GPU or CPU only systems.

*

**batch-size**
*Number of examples in each training batch (32)*

**clones**
*Number of model clones (calculated)

This value is automatically set to the number of available GPUs if `auto-
scale` is 'yes'.

When `auto-scale` is 'no' this value can be increased from 1 to train the
model in parallel on multiple GPUs.

*

**learning-rate**
*Initial learning rate (0.001)*

**learning-rate-decay-epochs**
*Number of epochs after which learning rate decays (2.0)*

**learning-rate-decay-factor**
*Learning rate decay factor (0.94)*

**learning-rate-decay-type**
*Method used to decay the learning rate (exponential)

Choices:
exponential
fixed
polynomial

*

**learning-rate-end**
*Minimal learning rate used by polynomial learning rate decay (0.0001)*

**log-save-seconds**
*Frequency of log summary saves in seconds (60)*

**log-steps**
*Frequency of summary logs in steps (100)*

**model-save-seconds**
*Frequency of model saves (checkpoints) in seconds (600)*

**optimizer**
*Optimizer used to train (rmsprop)

Choices:
adadelta
adagrad
adam
ftrl
momentum
rmsprop
sgd

*

**preprocessing**
*Preprocessing to use (default for model)*

**preprocessors**
*Number of preprocessing threads (calculated)

This value is automatically set to logical CPU count / 2 if `auto-scale`
is 'yes'.

When `auto-scale` is 'no' this value can be set to optimize the
preprocessor thread count for the system.

*

**readers**
*Number of parallel data readers (calculated)

This value is automatically set to logical CPU count / 2 if `auto-scale`
is 'yes'.

When `auto-scale` is 'no' this value can be set to optimize data reader
performance for the system.

*

**train-steps**
*Number of steps to train (train indefinitely)*

**weight-decay**
*Decay on the model weights (4e-05)*


pnasnet-large
=============

*TF-Slim PNASNet classifier*

Operations
^^^^^^^^^^

evaluate
--------

*Evaluate a trained model*

Flags
`````

**batch-size**
*Number of examples in each evaluated batch (100)*

**eval-batches**
*Number of batches to evaluate (all available)*

**step**
*Checkpoint step to evaluate (latest checkpoint)*

export-and-freeze
-----------------

*Export an inference graph with checkpoint weights*

Flags
`````

**step**
*Checkpoint step to use for the frozen graph (latest checkpoint)*

finetune
--------

*Finetune a trained model*

Flags
`````

**auto-scale**
*Adjust applicable flags for multi-GPU systems (yes)

Set to 'no' to disable any flag value adjustments.

When this value is 'yes' (the default) the following flags are adjusted on
multi-GPU systems:

- clones

- learning-rate

`clones` is set to the number of available GPUs.

`learning-rate` is adjusted by multiplying its specified value by the
number of GPUs.

Flags are not adjusted on single GPU or CPU only systems.

*

**batch-size**
*Number of examples in each training batch (32)*

**clones**
*Number of model clones (calculated)

This value is automatically set to the number of available GPUs if `auto-
scale` is 'yes'.

When `auto-scale` is 'no' this value can be increased from 1 to train the
model in parallel on multiple GPUs.

*

**learning-rate**
*Initial learning rate (0.0001)*

**learning-rate-decay-epochs**
*Number of epochs after which learning rate decays (2.0)*

**learning-rate-decay-factor**
*Learning rate decay factor (0.94)*

**learning-rate-decay-type**
*Method used to decay the learning rate (exponential)

Choices:
exponential
fixed
polynomial

*

**learning-rate-end**
*Minimal learning rate used by polynomial learning rate decay (0.0001)*

**log-save-seconds**
*Frequency of log summary saves in seconds (60)*

**log-steps**
*Frequency of summary logs in steps (100)*

**model-save-seconds**
*Frequency of model saves (checkpoints) in seconds (600)*

**optimizer**
*Optimizer used to train (rmsprop)

Choices:
adadelta
adagrad
adam
ftrl
momentum
rmsprop
sgd

*

**preprocessing**
*Preprocessing to use (default for model)*

**preprocessors**
*Number of preprocessing threads (calculated)

This value is automatically set to logical CPU count / 2 if `auto-scale`
is 'yes'.

When `auto-scale` is 'no' this value can be set to optimize the
preprocessor thread count for the system.

*

**readers**
*Number of parallel data readers (calculated)

This value is automatically set to logical CPU count / 2 if `auto-scale`
is 'yes'.

When `auto-scale` is 'no' this value can be set to optimize data reader
performance for the system.

*

**train-steps**
*Number of steps to train (train indefinitely)*

**weight-decay**
*Decay on the model weights (4e-05)*

label
-----

*Classify an image using a trained model*

Flags
`````

**image**
*Path to image to classify (required)*

tflite
------

*Generate a TFLite file from a frozen graph*

Flags
`````

**output-format**
*TF Lite output format (tflite)

Choices:
tflite
graphviz_dot

*

**quantized**
*Whether or not output arrays are quantized (no)

Choices:
yes
no

*

**quantized-inputs**
*Whether or not input arrays are quantized (no)

Choices:
yes
no

*

train
-----

*Train model from scratch*

Flags
`````

**auto-scale**
*Adjust applicable flags for multi-GPU systems (yes)

Set to 'no' to disable any flag value adjustments.

When this value is 'yes' (the default) the following flags are adjusted on
multi-GPU systems:

- clones

- learning-rate

`clones` is set to the number of available GPUs.

`learning-rate` is adjusted by multiplying its specified value by the
number of GPUs.

Flags are not adjusted on single GPU or CPU only systems.

*

**batch-size**
*Number of examples in each training batch (32)*

**clones**
*Number of model clones (calculated)

This value is automatically set to the number of available GPUs if `auto-
scale` is 'yes'.

When `auto-scale` is 'no' this value can be increased from 1 to train the
model in parallel on multiple GPUs.

*

**learning-rate**
*Initial learning rate (0.001)*

**learning-rate-decay-epochs**
*Number of epochs after which learning rate decays (2.0)*

**learning-rate-decay-factor**
*Learning rate decay factor (0.94)*

**learning-rate-decay-type**
*Method used to decay the learning rate (exponential)

Choices:
exponential
fixed
polynomial

*

**learning-rate-end**
*Minimal learning rate used by polynomial learning rate decay (0.0001)*

**log-save-seconds**
*Frequency of log summary saves in seconds (60)*

**log-steps**
*Frequency of summary logs in steps (100)*

**model-save-seconds**
*Frequency of model saves (checkpoints) in seconds (600)*

**optimizer**
*Optimizer used to train (rmsprop)

Choices:
adadelta
adagrad
adam
ftrl
momentum
rmsprop
sgd

*

**preprocessing**
*Preprocessing to use (default for model)*

**preprocessors**
*Number of preprocessing threads (calculated)

This value is automatically set to logical CPU count / 2 if `auto-scale`
is 'yes'.

When `auto-scale` is 'no' this value can be set to optimize the
preprocessor thread count for the system.

*

**readers**
*Number of parallel data readers (calculated)

This value is automatically set to logical CPU count / 2 if `auto-scale`
is 'yes'.

When `auto-scale` is 'no' this value can be set to optimize data reader
performance for the system.

*

**train-steps**
*Number of steps to train (train indefinitely)*

**weight-decay**
*Decay on the model weights (4e-05)*

transfer-learn
--------------

*Train model using transfer learning*

Flags
`````

**auto-scale**
*Adjust applicable flags for multi-GPU systems (yes)

Set to 'no' to disable any flag value adjustments.

When this value is 'yes' (the default) the following flags are adjusted on
multi-GPU systems:

- clones

- learning-rate

`clones` is set to the number of available GPUs.

`learning-rate` is adjusted by multiplying its specified value by the
number of GPUs.

Flags are not adjusted on single GPU or CPU only systems.

*

**batch-size**
*Number of examples in each training batch (32)*

**clones**
*Number of model clones (calculated)

This value is automatically set to the number of available GPUs if `auto-
scale` is 'yes'.

When `auto-scale` is 'no' this value can be increased from 1 to train the
model in parallel on multiple GPUs.

*

**learning-rate**
*Initial learning rate (0.001)*

**learning-rate-decay-epochs**
*Number of epochs after which learning rate decays (2.0)*

**learning-rate-decay-factor**
*Learning rate decay factor (0.94)*

**learning-rate-decay-type**
*Method used to decay the learning rate (exponential)

Choices:
exponential
fixed
polynomial

*

**learning-rate-end**
*Minimal learning rate used by polynomial learning rate decay (0.0001)*

**log-save-seconds**
*Frequency of log summary saves in seconds (60)*

**log-steps**
*Frequency of summary logs in steps (100)*

**model-save-seconds**
*Frequency of model saves (checkpoints) in seconds (600)*

**optimizer**
*Optimizer used to train (rmsprop)

Choices:
adadelta
adagrad
adam
ftrl
momentum
rmsprop
sgd

*

**preprocessing**
*Preprocessing to use (default for model)*

**preprocessors**
*Number of preprocessing threads (calculated)

This value is automatically set to logical CPU count / 2 if `auto-scale`
is 'yes'.

When `auto-scale` is 'no' this value can be set to optimize the
preprocessor thread count for the system.

*

**readers**
*Number of parallel data readers (calculated)

This value is automatically set to logical CPU count / 2 if `auto-scale`
is 'yes'.

When `auto-scale` is 'no' this value can be set to optimize data reader
performance for the system.

*

**train-steps**
*Number of steps to train (train indefinitely)*

**weight-decay**
*Decay on the model weights (4e-05)*


pnasnet-mobile
==============

*TF-Slim PNASNet mobile classifier*

Operations
^^^^^^^^^^

evaluate
--------

*Evaluate a trained model*

Flags
`````

**batch-size**
*Number of examples in each evaluated batch (100)*

**eval-batches**
*Number of batches to evaluate (all available)*

**step**
*Checkpoint step to evaluate (latest checkpoint)*

export-and-freeze
-----------------

*Export an inference graph with checkpoint weights*

Flags
`````

**step**
*Checkpoint step to use for the frozen graph (latest checkpoint)*

finetune
--------

*Finetune a trained model*

Flags
`````

**auto-scale**
*Adjust applicable flags for multi-GPU systems (yes)

Set to 'no' to disable any flag value adjustments.

When this value is 'yes' (the default) the following flags are adjusted on
multi-GPU systems:

- clones

- learning-rate

`clones` is set to the number of available GPUs.

`learning-rate` is adjusted by multiplying its specified value by the
number of GPUs.

Flags are not adjusted on single GPU or CPU only systems.

*

**batch-size**
*Number of examples in each training batch (32)*

**clones**
*Number of model clones (calculated)

This value is automatically set to the number of available GPUs if `auto-
scale` is 'yes'.

When `auto-scale` is 'no' this value can be increased from 1 to train the
model in parallel on multiple GPUs.

*

**learning-rate**
*Initial learning rate (0.0001)*

**learning-rate-decay-epochs**
*Number of epochs after which learning rate decays (2.0)*

**learning-rate-decay-factor**
*Learning rate decay factor (0.94)*

**learning-rate-decay-type**
*Method used to decay the learning rate (exponential)

Choices:
exponential
fixed
polynomial

*

**learning-rate-end**
*Minimal learning rate used by polynomial learning rate decay (0.0001)*

**log-save-seconds**
*Frequency of log summary saves in seconds (60)*

**log-steps**
*Frequency of summary logs in steps (100)*

**model-save-seconds**
*Frequency of model saves (checkpoints) in seconds (600)*

**optimizer**
*Optimizer used to train (rmsprop)

Choices:
adadelta
adagrad
adam
ftrl
momentum
rmsprop
sgd

*

**preprocessing**
*Preprocessing to use (default for model)*

**preprocessors**
*Number of preprocessing threads (calculated)

This value is automatically set to logical CPU count / 2 if `auto-scale`
is 'yes'.

When `auto-scale` is 'no' this value can be set to optimize the
preprocessor thread count for the system.

*

**readers**
*Number of parallel data readers (calculated)

This value is automatically set to logical CPU count / 2 if `auto-scale`
is 'yes'.

When `auto-scale` is 'no' this value can be set to optimize data reader
performance for the system.

*

**train-steps**
*Number of steps to train (train indefinitely)*

**weight-decay**
*Decay on the model weights (4e-05)*

label
-----

*Classify an image using a trained model*

Flags
`````

**image**
*Path to image to classify (required)*

tflite
------

*Generate a TFLite file from a frozen graph*

Flags
`````

**output-format**
*TF Lite output format (tflite)

Choices:
tflite
graphviz_dot

*

**quantized**
*Whether or not output arrays are quantized (no)

Choices:
yes
no

*

**quantized-inputs**
*Whether or not input arrays are quantized (no)

Choices:
yes
no

*

train
-----

*Train model from scratch*

Flags
`````

**auto-scale**
*Adjust applicable flags for multi-GPU systems (yes)

Set to 'no' to disable any flag value adjustments.

When this value is 'yes' (the default) the following flags are adjusted on
multi-GPU systems:

- clones

- learning-rate

`clones` is set to the number of available GPUs.

`learning-rate` is adjusted by multiplying its specified value by the
number of GPUs.

Flags are not adjusted on single GPU or CPU only systems.

*

**batch-size**
*Number of examples in each training batch (32)*

**clones**
*Number of model clones (calculated)

This value is automatically set to the number of available GPUs if `auto-
scale` is 'yes'.

When `auto-scale` is 'no' this value can be increased from 1 to train the
model in parallel on multiple GPUs.

*

**learning-rate**
*Initial learning rate (0.001)*

**learning-rate-decay-epochs**
*Number of epochs after which learning rate decays (2.0)*

**learning-rate-decay-factor**
*Learning rate decay factor (0.94)*

**learning-rate-decay-type**
*Method used to decay the learning rate (exponential)

Choices:
exponential
fixed
polynomial

*

**learning-rate-end**
*Minimal learning rate used by polynomial learning rate decay (0.0001)*

**log-save-seconds**
*Frequency of log summary saves in seconds (60)*

**log-steps**
*Frequency of summary logs in steps (100)*

**model-save-seconds**
*Frequency of model saves (checkpoints) in seconds (600)*

**optimizer**
*Optimizer used to train (rmsprop)

Choices:
adadelta
adagrad
adam
ftrl
momentum
rmsprop
sgd

*

**preprocessing**
*Preprocessing to use (default for model)*

**preprocessors**
*Number of preprocessing threads (calculated)

This value is automatically set to logical CPU count / 2 if `auto-scale`
is 'yes'.

When `auto-scale` is 'no' this value can be set to optimize the
preprocessor thread count for the system.

*

**readers**
*Number of parallel data readers (calculated)

This value is automatically set to logical CPU count / 2 if `auto-scale`
is 'yes'.

When `auto-scale` is 'no' this value can be set to optimize data reader
performance for the system.

*

**train-steps**
*Number of steps to train (train indefinitely)*

**weight-decay**
*Decay on the model weights (4e-05)*

transfer-learn
--------------

*Train model using transfer learning*

Flags
`````

**auto-scale**
*Adjust applicable flags for multi-GPU systems (yes)

Set to 'no' to disable any flag value adjustments.

When this value is 'yes' (the default) the following flags are adjusted on
multi-GPU systems:

- clones

- learning-rate

`clones` is set to the number of available GPUs.

`learning-rate` is adjusted by multiplying its specified value by the
number of GPUs.

Flags are not adjusted on single GPU or CPU only systems.

*

**batch-size**
*Number of examples in each training batch (32)*

**clones**
*Number of model clones (calculated)

This value is automatically set to the number of available GPUs if `auto-
scale` is 'yes'.

When `auto-scale` is 'no' this value can be increased from 1 to train the
model in parallel on multiple GPUs.

*

**learning-rate**
*Initial learning rate (0.001)*

**learning-rate-decay-epochs**
*Number of epochs after which learning rate decays (2.0)*

**learning-rate-decay-factor**
*Learning rate decay factor (0.94)*

**learning-rate-decay-type**
*Method used to decay the learning rate (exponential)

Choices:
exponential
fixed
polynomial

*

**learning-rate-end**
*Minimal learning rate used by polynomial learning rate decay (0.0001)*

**log-save-seconds**
*Frequency of log summary saves in seconds (60)*

**log-steps**
*Frequency of summary logs in steps (100)*

**model-save-seconds**
*Frequency of model saves (checkpoints) in seconds (600)*

**optimizer**
*Optimizer used to train (rmsprop)

Choices:
adadelta
adagrad
adam
ftrl
momentum
rmsprop
sgd

*

**preprocessing**
*Preprocessing to use (default for model)*

**preprocessors**
*Number of preprocessing threads (calculated)

This value is automatically set to logical CPU count / 2 if `auto-scale`
is 'yes'.

When `auto-scale` is 'no' this value can be set to optimize the
preprocessor thread count for the system.

*

**readers**
*Number of parallel data readers (calculated)

This value is automatically set to logical CPU count / 2 if `auto-scale`
is 'yes'.

When `auto-scale` is 'no' this value can be set to optimize data reader
performance for the system.

*

**train-steps**
*Number of steps to train (train indefinitely)*

**weight-decay**
*Decay on the model weights (4e-05)*


resnet-101
==========

*TF-Slim ResNet v1 101 layer classifier*

Operations
^^^^^^^^^^

evaluate
--------

*Evaluate a trained model*

Flags
`````

**batch-size**
*Number of examples in each evaluated batch (100)*

**eval-batches**
*Number of batches to evaluate (all available)*

**step**
*Checkpoint step to evaluate (latest checkpoint)*

export-and-freeze
-----------------

*Export an inference graph with checkpoint weights*

Flags
`````

**step**
*Checkpoint step to use for the frozen graph (latest checkpoint)*

finetune
--------

*Finetune a trained model*

Flags
`````

**auto-scale**
*Adjust applicable flags for multi-GPU systems (yes)

Set to 'no' to disable any flag value adjustments.

When this value is 'yes' (the default) the following flags are adjusted on
multi-GPU systems:

- clones

- learning-rate

`clones` is set to the number of available GPUs.

`learning-rate` is adjusted by multiplying its specified value by the
number of GPUs.

Flags are not adjusted on single GPU or CPU only systems.

*

**batch-size**
*Number of examples in each training batch (32)*

**clones**
*Number of model clones (calculated)

This value is automatically set to the number of available GPUs if `auto-
scale` is 'yes'.

When `auto-scale` is 'no' this value can be increased from 1 to train the
model in parallel on multiple GPUs.

*

**learning-rate**
*Initial learning rate (0.0001)*

**learning-rate-decay-epochs**
*Number of epochs after which learning rate decays (2.0)*

**learning-rate-decay-factor**
*Learning rate decay factor (0.94)*

**learning-rate-decay-type**
*Method used to decay the learning rate (exponential)

Choices:
exponential
fixed
polynomial

*

**learning-rate-end**
*Minimal learning rate used by polynomial learning rate decay (0.0001)*

**log-save-seconds**
*Frequency of log summary saves in seconds (60)*

**log-steps**
*Frequency of summary logs in steps (100)*

**model-save-seconds**
*Frequency of model saves (checkpoints) in seconds (600)*

**optimizer**
*Optimizer used to train (rmsprop)

Choices:
adadelta
adagrad
adam
ftrl
momentum
rmsprop
sgd

*

**preprocessing**
*Preprocessing to use (default for model)*

**preprocessors**
*Number of preprocessing threads (calculated)

This value is automatically set to logical CPU count / 2 if `auto-scale`
is 'yes'.

When `auto-scale` is 'no' this value can be set to optimize the
preprocessor thread count for the system.

*

**readers**
*Number of parallel data readers (calculated)

This value is automatically set to logical CPU count / 2 if `auto-scale`
is 'yes'.

When `auto-scale` is 'no' this value can be set to optimize data reader
performance for the system.

*

**train-steps**
*Number of steps to train (train indefinitely)*

**weight-decay**
*Decay on the model weights (4e-05)*

label
-----

*Classify an image using a trained model*

Flags
`````

**image**
*Path to image to classify (required)*

tflite
------

*Generate a TFLite file from a frozen graph*

Flags
`````

**output-format**
*TF Lite output format (tflite)

Choices:
tflite
graphviz_dot

*

**quantized**
*Whether or not output arrays are quantized (no)

Choices:
yes
no

*

**quantized-inputs**
*Whether or not input arrays are quantized (no)

Choices:
yes
no

*

train
-----

*Train model from scratch*

Flags
`````

**auto-scale**
*Adjust applicable flags for multi-GPU systems (yes)

Set to 'no' to disable any flag value adjustments.

When this value is 'yes' (the default) the following flags are adjusted on
multi-GPU systems:

- clones

- learning-rate

`clones` is set to the number of available GPUs.

`learning-rate` is adjusted by multiplying its specified value by the
number of GPUs.

Flags are not adjusted on single GPU or CPU only systems.

*

**batch-size**
*Number of examples in each training batch (32)*

**clones**
*Number of model clones (calculated)

This value is automatically set to the number of available GPUs if `auto-
scale` is 'yes'.

When `auto-scale` is 'no' this value can be increased from 1 to train the
model in parallel on multiple GPUs.

*

**learning-rate**
*Initial learning rate (0.001)*

**learning-rate-decay-epochs**
*Number of epochs after which learning rate decays (2.0)*

**learning-rate-decay-factor**
*Learning rate decay factor (0.94)*

**learning-rate-decay-type**
*Method used to decay the learning rate (exponential)

Choices:
exponential
fixed
polynomial

*

**learning-rate-end**
*Minimal learning rate used by polynomial learning rate decay (0.0001)*

**log-save-seconds**
*Frequency of log summary saves in seconds (60)*

**log-steps**
*Frequency of summary logs in steps (100)*

**model-save-seconds**
*Frequency of model saves (checkpoints) in seconds (600)*

**optimizer**
*Optimizer used to train (rmsprop)

Choices:
adadelta
adagrad
adam
ftrl
momentum
rmsprop
sgd

*

**preprocessing**
*Preprocessing to use (default for model)*

**preprocessors**
*Number of preprocessing threads (calculated)

This value is automatically set to logical CPU count / 2 if `auto-scale`
is 'yes'.

When `auto-scale` is 'no' this value can be set to optimize the
preprocessor thread count for the system.

*

**readers**
*Number of parallel data readers (calculated)

This value is automatically set to logical CPU count / 2 if `auto-scale`
is 'yes'.

When `auto-scale` is 'no' this value can be set to optimize data reader
performance for the system.

*

**train-steps**
*Number of steps to train (train indefinitely)*

**weight-decay**
*Decay on the model weights (4e-05)*

transfer-learn
--------------

*Train model using transfer learning*

Flags
`````

**auto-scale**
*Adjust applicable flags for multi-GPU systems (yes)

Set to 'no' to disable any flag value adjustments.

When this value is 'yes' (the default) the following flags are adjusted on
multi-GPU systems:

- clones

- learning-rate

`clones` is set to the number of available GPUs.

`learning-rate` is adjusted by multiplying its specified value by the
number of GPUs.

Flags are not adjusted on single GPU or CPU only systems.

*

**batch-size**
*Number of examples in each training batch (32)*

**clones**
*Number of model clones (calculated)

This value is automatically set to the number of available GPUs if `auto-
scale` is 'yes'.

When `auto-scale` is 'no' this value can be increased from 1 to train the
model in parallel on multiple GPUs.

*

**learning-rate**
*Initial learning rate (0.001)*

**learning-rate-decay-epochs**
*Number of epochs after which learning rate decays (2.0)*

**learning-rate-decay-factor**
*Learning rate decay factor (0.94)*

**learning-rate-decay-type**
*Method used to decay the learning rate (exponential)

Choices:
exponential
fixed
polynomial

*

**learning-rate-end**
*Minimal learning rate used by polynomial learning rate decay (0.0001)*

**log-save-seconds**
*Frequency of log summary saves in seconds (60)*

**log-steps**
*Frequency of summary logs in steps (100)*

**model-save-seconds**
*Frequency of model saves (checkpoints) in seconds (600)*

**optimizer**
*Optimizer used to train (rmsprop)

Choices:
adadelta
adagrad
adam
ftrl
momentum
rmsprop
sgd

*

**preprocessing**
*Preprocessing to use (default for model)*

**preprocessors**
*Number of preprocessing threads (calculated)

This value is automatically set to logical CPU count / 2 if `auto-scale`
is 'yes'.

When `auto-scale` is 'no' this value can be set to optimize the
preprocessor thread count for the system.

*

**readers**
*Number of parallel data readers (calculated)

This value is automatically set to logical CPU count / 2 if `auto-scale`
is 'yes'.

When `auto-scale` is 'no' this value can be set to optimize data reader
performance for the system.

*

**train-steps**
*Number of steps to train (train indefinitely)*

**weight-decay**
*Decay on the model weights (4e-05)*


resnet-152
==========

*TF-Slim ResNet v1 152 layer classifier*

Operations
^^^^^^^^^^

evaluate
--------

*Evaluate a trained model*

Flags
`````

**batch-size**
*Number of examples in each evaluated batch (100)*

**eval-batches**
*Number of batches to evaluate (all available)*

**step**
*Checkpoint step to evaluate (latest checkpoint)*

export-and-freeze
-----------------

*Export an inference graph with checkpoint weights*

Flags
`````

**step**
*Checkpoint step to use for the frozen graph (latest checkpoint)*

finetune
--------

*Finetune a trained model*

Flags
`````

**auto-scale**
*Adjust applicable flags for multi-GPU systems (yes)

Set to 'no' to disable any flag value adjustments.

When this value is 'yes' (the default) the following flags are adjusted on
multi-GPU systems:

- clones

- learning-rate

`clones` is set to the number of available GPUs.

`learning-rate` is adjusted by multiplying its specified value by the
number of GPUs.

Flags are not adjusted on single GPU or CPU only systems.

*

**batch-size**
*Number of examples in each training batch (32)*

**clones**
*Number of model clones (calculated)

This value is automatically set to the number of available GPUs if `auto-
scale` is 'yes'.

When `auto-scale` is 'no' this value can be increased from 1 to train the
model in parallel on multiple GPUs.

*

**learning-rate**
*Initial learning rate (0.0001)*

**learning-rate-decay-epochs**
*Number of epochs after which learning rate decays (2.0)*

**learning-rate-decay-factor**
*Learning rate decay factor (0.94)*

**learning-rate-decay-type**
*Method used to decay the learning rate (exponential)

Choices:
exponential
fixed
polynomial

*

**learning-rate-end**
*Minimal learning rate used by polynomial learning rate decay (0.0001)*

**log-save-seconds**
*Frequency of log summary saves in seconds (60)*

**log-steps**
*Frequency of summary logs in steps (100)*

**model-save-seconds**
*Frequency of model saves (checkpoints) in seconds (600)*

**optimizer**
*Optimizer used to train (rmsprop)

Choices:
adadelta
adagrad
adam
ftrl
momentum
rmsprop
sgd

*

**preprocessing**
*Preprocessing to use (default for model)*

**preprocessors**
*Number of preprocessing threads (calculated)

This value is automatically set to logical CPU count / 2 if `auto-scale`
is 'yes'.

When `auto-scale` is 'no' this value can be set to optimize the
preprocessor thread count for the system.

*

**readers**
*Number of parallel data readers (calculated)

This value is automatically set to logical CPU count / 2 if `auto-scale`
is 'yes'.

When `auto-scale` is 'no' this value can be set to optimize data reader
performance for the system.

*

**train-steps**
*Number of steps to train (train indefinitely)*

**weight-decay**
*Decay on the model weights (4e-05)*

label
-----

*Classify an image using a trained model*

Flags
`````

**image**
*Path to image to classify (required)*

tflite
------

*Generate a TFLite file from a frozen graph*

Flags
`````

**output-format**
*TF Lite output format (tflite)

Choices:
tflite
graphviz_dot

*

**quantized**
*Whether or not output arrays are quantized (no)

Choices:
yes
no

*

**quantized-inputs**
*Whether or not input arrays are quantized (no)

Choices:
yes
no

*

train
-----

*Train model from scratch*

Flags
`````

**auto-scale**
*Adjust applicable flags for multi-GPU systems (yes)

Set to 'no' to disable any flag value adjustments.

When this value is 'yes' (the default) the following flags are adjusted on
multi-GPU systems:

- clones

- learning-rate

`clones` is set to the number of available GPUs.

`learning-rate` is adjusted by multiplying its specified value by the
number of GPUs.

Flags are not adjusted on single GPU or CPU only systems.

*

**batch-size**
*Number of examples in each training batch (32)*

**clones**
*Number of model clones (calculated)

This value is automatically set to the number of available GPUs if `auto-
scale` is 'yes'.

When `auto-scale` is 'no' this value can be increased from 1 to train the
model in parallel on multiple GPUs.

*

**learning-rate**
*Initial learning rate (0.001)*

**learning-rate-decay-epochs**
*Number of epochs after which learning rate decays (2.0)*

**learning-rate-decay-factor**
*Learning rate decay factor (0.94)*

**learning-rate-decay-type**
*Method used to decay the learning rate (exponential)

Choices:
exponential
fixed
polynomial

*

**learning-rate-end**
*Minimal learning rate used by polynomial learning rate decay (0.0001)*

**log-save-seconds**
*Frequency of log summary saves in seconds (60)*

**log-steps**
*Frequency of summary logs in steps (100)*

**model-save-seconds**
*Frequency of model saves (checkpoints) in seconds (600)*

**optimizer**
*Optimizer used to train (rmsprop)

Choices:
adadelta
adagrad
adam
ftrl
momentum
rmsprop
sgd

*

**preprocessing**
*Preprocessing to use (default for model)*

**preprocessors**
*Number of preprocessing threads (calculated)

This value is automatically set to logical CPU count / 2 if `auto-scale`
is 'yes'.

When `auto-scale` is 'no' this value can be set to optimize the
preprocessor thread count for the system.

*

**readers**
*Number of parallel data readers (calculated)

This value is automatically set to logical CPU count / 2 if `auto-scale`
is 'yes'.

When `auto-scale` is 'no' this value can be set to optimize data reader
performance for the system.

*

**train-steps**
*Number of steps to train (train indefinitely)*

**weight-decay**
*Decay on the model weights (4e-05)*

transfer-learn
--------------

*Train model using transfer learning*

Flags
`````

**auto-scale**
*Adjust applicable flags for multi-GPU systems (yes)

Set to 'no' to disable any flag value adjustments.

When this value is 'yes' (the default) the following flags are adjusted on
multi-GPU systems:

- clones

- learning-rate

`clones` is set to the number of available GPUs.

`learning-rate` is adjusted by multiplying its specified value by the
number of GPUs.

Flags are not adjusted on single GPU or CPU only systems.

*

**batch-size**
*Number of examples in each training batch (32)*

**clones**
*Number of model clones (calculated)

This value is automatically set to the number of available GPUs if `auto-
scale` is 'yes'.

When `auto-scale` is 'no' this value can be increased from 1 to train the
model in parallel on multiple GPUs.

*

**learning-rate**
*Initial learning rate (0.001)*

**learning-rate-decay-epochs**
*Number of epochs after which learning rate decays (2.0)*

**learning-rate-decay-factor**
*Learning rate decay factor (0.94)*

**learning-rate-decay-type**
*Method used to decay the learning rate (exponential)

Choices:
exponential
fixed
polynomial

*

**learning-rate-end**
*Minimal learning rate used by polynomial learning rate decay (0.0001)*

**log-save-seconds**
*Frequency of log summary saves in seconds (60)*

**log-steps**
*Frequency of summary logs in steps (100)*

**model-save-seconds**
*Frequency of model saves (checkpoints) in seconds (600)*

**optimizer**
*Optimizer used to train (rmsprop)

Choices:
adadelta
adagrad
adam
ftrl
momentum
rmsprop
sgd

*

**preprocessing**
*Preprocessing to use (default for model)*

**preprocessors**
*Number of preprocessing threads (calculated)

This value is automatically set to logical CPU count / 2 if `auto-scale`
is 'yes'.

When `auto-scale` is 'no' this value can be set to optimize the
preprocessor thread count for the system.

*

**readers**
*Number of parallel data readers (calculated)

This value is automatically set to logical CPU count / 2 if `auto-scale`
is 'yes'.

When `auto-scale` is 'no' this value can be set to optimize data reader
performance for the system.

*

**train-steps**
*Number of steps to train (train indefinitely)*

**weight-decay**
*Decay on the model weights (4e-05)*


resnet-50
=========

*TF-Slim ResNet v1 50 layer classifier*

Operations
^^^^^^^^^^

evaluate
--------

*Evaluate a trained model*

Flags
`````

**batch-size**
*Number of examples in each evaluated batch (100)*

**eval-batches**
*Number of batches to evaluate (all available)*

**step**
*Checkpoint step to evaluate (latest checkpoint)*

export-and-freeze
-----------------

*Export an inference graph with checkpoint weights*

Flags
`````

**step**
*Checkpoint step to use for the frozen graph (latest checkpoint)*

finetune
--------

*Finetune a trained model*

Flags
`````

**auto-scale**
*Adjust applicable flags for multi-GPU systems (yes)

Set to 'no' to disable any flag value adjustments.

When this value is 'yes' (the default) the following flags are adjusted on
multi-GPU systems:

- clones

- learning-rate

`clones` is set to the number of available GPUs.

`learning-rate` is adjusted by multiplying its specified value by the
number of GPUs.

Flags are not adjusted on single GPU or CPU only systems.

*

**batch-size**
*Number of examples in each training batch (32)*

**clones**
*Number of model clones (calculated)

This value is automatically set to the number of available GPUs if `auto-
scale` is 'yes'.

When `auto-scale` is 'no' this value can be increased from 1 to train the
model in parallel on multiple GPUs.

*

**learning-rate**
*Initial learning rate (0.0001)*

**learning-rate-decay-epochs**
*Number of epochs after which learning rate decays (2.0)*

**learning-rate-decay-factor**
*Learning rate decay factor (0.94)*

**learning-rate-decay-type**
*Method used to decay the learning rate (exponential)

Choices:
exponential
fixed
polynomial

*

**learning-rate-end**
*Minimal learning rate used by polynomial learning rate decay (0.0001)*

**log-save-seconds**
*Frequency of log summary saves in seconds (60)*

**log-steps**
*Frequency of summary logs in steps (100)*

**model-save-seconds**
*Frequency of model saves (checkpoints) in seconds (600)*

**optimizer**
*Optimizer used to train (rmsprop)

Choices:
adadelta
adagrad
adam
ftrl
momentum
rmsprop
sgd

*

**preprocessing**
*Preprocessing to use (default for model)*

**preprocessors**
*Number of preprocessing threads (calculated)

This value is automatically set to logical CPU count / 2 if `auto-scale`
is 'yes'.

When `auto-scale` is 'no' this value can be set to optimize the
preprocessor thread count for the system.

*

**readers**
*Number of parallel data readers (calculated)

This value is automatically set to logical CPU count / 2 if `auto-scale`
is 'yes'.

When `auto-scale` is 'no' this value can be set to optimize data reader
performance for the system.

*

**train-steps**
*Number of steps to train (train indefinitely)*

**weight-decay**
*Decay on the model weights (4e-05)*

label
-----

*Classify an image using a trained model*

Flags
`````

**image**
*Path to image to classify (required)*

tflite
------

*Generate a TFLite file from a frozen graph*

Flags
`````

**output-format**
*TF Lite output format (tflite)

Choices:
tflite
graphviz_dot

*

**quantized**
*Whether or not output arrays are quantized (no)

Choices:
yes
no

*

**quantized-inputs**
*Whether or not input arrays are quantized (no)

Choices:
yes
no

*

train
-----

*Train model from scratch*

Flags
`````

**auto-scale**
*Adjust applicable flags for multi-GPU systems (yes)

Set to 'no' to disable any flag value adjustments.

When this value is 'yes' (the default) the following flags are adjusted on
multi-GPU systems:

- clones

- learning-rate

`clones` is set to the number of available GPUs.

`learning-rate` is adjusted by multiplying its specified value by the
number of GPUs.

Flags are not adjusted on single GPU or CPU only systems.

*

**batch-size**
*Number of examples in each training batch (32)*

**clones**
*Number of model clones (calculated)

This value is automatically set to the number of available GPUs if `auto-
scale` is 'yes'.

When `auto-scale` is 'no' this value can be increased from 1 to train the
model in parallel on multiple GPUs.

*

**learning-rate**
*Initial learning rate (0.001)*

**learning-rate-decay-epochs**
*Number of epochs after which learning rate decays (2.0)*

**learning-rate-decay-factor**
*Learning rate decay factor (0.94)*

**learning-rate-decay-type**
*Method used to decay the learning rate (exponential)

Choices:
exponential
fixed
polynomial

*

**learning-rate-end**
*Minimal learning rate used by polynomial learning rate decay (0.0001)*

**log-save-seconds**
*Frequency of log summary saves in seconds (60)*

**log-steps**
*Frequency of summary logs in steps (100)*

**model-save-seconds**
*Frequency of model saves (checkpoints) in seconds (600)*

**optimizer**
*Optimizer used to train (rmsprop)

Choices:
adadelta
adagrad
adam
ftrl
momentum
rmsprop
sgd

*

**preprocessing**
*Preprocessing to use (default for model)*

**preprocessors**
*Number of preprocessing threads (calculated)

This value is automatically set to logical CPU count / 2 if `auto-scale`
is 'yes'.

When `auto-scale` is 'no' this value can be set to optimize the
preprocessor thread count for the system.

*

**readers**
*Number of parallel data readers (calculated)

This value is automatically set to logical CPU count / 2 if `auto-scale`
is 'yes'.

When `auto-scale` is 'no' this value can be set to optimize data reader
performance for the system.

*

**train-steps**
*Number of steps to train (train indefinitely)*

**weight-decay**
*Decay on the model weights (4e-05)*

transfer-learn
--------------

*Train model using transfer learning*

Flags
`````

**auto-scale**
*Adjust applicable flags for multi-GPU systems (yes)

Set to 'no' to disable any flag value adjustments.

When this value is 'yes' (the default) the following flags are adjusted on
multi-GPU systems:

- clones

- learning-rate

`clones` is set to the number of available GPUs.

`learning-rate` is adjusted by multiplying its specified value by the
number of GPUs.

Flags are not adjusted on single GPU or CPU only systems.

*

**batch-size**
*Number of examples in each training batch (32)*

**clones**
*Number of model clones (calculated)

This value is automatically set to the number of available GPUs if `auto-
scale` is 'yes'.

When `auto-scale` is 'no' this value can be increased from 1 to train the
model in parallel on multiple GPUs.

*

**learning-rate**
*Initial learning rate (0.001)*

**learning-rate-decay-epochs**
*Number of epochs after which learning rate decays (2.0)*

**learning-rate-decay-factor**
*Learning rate decay factor (0.94)*

**learning-rate-decay-type**
*Method used to decay the learning rate (exponential)

Choices:
exponential
fixed
polynomial

*

**learning-rate-end**
*Minimal learning rate used by polynomial learning rate decay (0.0001)*

**log-save-seconds**
*Frequency of log summary saves in seconds (60)*

**log-steps**
*Frequency of summary logs in steps (100)*

**model-save-seconds**
*Frequency of model saves (checkpoints) in seconds (600)*

**optimizer**
*Optimizer used to train (rmsprop)

Choices:
adadelta
adagrad
adam
ftrl
momentum
rmsprop
sgd

*

**preprocessing**
*Preprocessing to use (default for model)*

**preprocessors**
*Number of preprocessing threads (calculated)

This value is automatically set to logical CPU count / 2 if `auto-scale`
is 'yes'.

When `auto-scale` is 'no' this value can be set to optimize the
preprocessor thread count for the system.

*

**readers**
*Number of parallel data readers (calculated)

This value is automatically set to logical CPU count / 2 if `auto-scale`
is 'yes'.

When `auto-scale` is 'no' this value can be set to optimize data reader
performance for the system.

*

**train-steps**
*Number of steps to train (train indefinitely)*

**weight-decay**
*Decay on the model weights (4e-05)*


resnet-v2-101
=============

*TF-Slim ResNet v2 101 layer classifier*

Operations
^^^^^^^^^^

evaluate
--------

*Evaluate a trained model*

Flags
`````

**batch-size**
*Number of examples in each evaluated batch (100)*

**eval-batches**
*Number of batches to evaluate (all available)*

**step**
*Checkpoint step to evaluate (latest checkpoint)*

export-and-freeze
-----------------

*Export an inference graph with checkpoint weights*

Flags
`````

**step**
*Checkpoint step to use for the frozen graph (latest checkpoint)*

finetune
--------

*Finetune a trained model*

Flags
`````

**auto-scale**
*Adjust applicable flags for multi-GPU systems (yes)

Set to 'no' to disable any flag value adjustments.

When this value is 'yes' (the default) the following flags are adjusted on
multi-GPU systems:

- clones

- learning-rate

`clones` is set to the number of available GPUs.

`learning-rate` is adjusted by multiplying its specified value by the
number of GPUs.

Flags are not adjusted on single GPU or CPU only systems.

*

**batch-size**
*Number of examples in each training batch (32)*

**clones**
*Number of model clones (calculated)

This value is automatically set to the number of available GPUs if `auto-
scale` is 'yes'.

When `auto-scale` is 'no' this value can be increased from 1 to train the
model in parallel on multiple GPUs.

*

**learning-rate**
*Initial learning rate (0.0001)*

**learning-rate-decay-epochs**
*Number of epochs after which learning rate decays (2.0)*

**learning-rate-decay-factor**
*Learning rate decay factor (0.94)*

**learning-rate-decay-type**
*Method used to decay the learning rate (exponential)

Choices:
exponential
fixed
polynomial

*

**learning-rate-end**
*Minimal learning rate used by polynomial learning rate decay (0.0001)*

**log-save-seconds**
*Frequency of log summary saves in seconds (60)*

**log-steps**
*Frequency of summary logs in steps (100)*

**model-save-seconds**
*Frequency of model saves (checkpoints) in seconds (600)*

**optimizer**
*Optimizer used to train (rmsprop)

Choices:
adadelta
adagrad
adam
ftrl
momentum
rmsprop
sgd

*

**preprocessing**
*Preprocessing to use (default for model)*

**preprocessors**
*Number of preprocessing threads (calculated)

This value is automatically set to logical CPU count / 2 if `auto-scale`
is 'yes'.

When `auto-scale` is 'no' this value can be set to optimize the
preprocessor thread count for the system.

*

**readers**
*Number of parallel data readers (calculated)

This value is automatically set to logical CPU count / 2 if `auto-scale`
is 'yes'.

When `auto-scale` is 'no' this value can be set to optimize data reader
performance for the system.

*

**train-steps**
*Number of steps to train (train indefinitely)*

**weight-decay**
*Decay on the model weights (4e-05)*

label
-----

*Classify an image using a trained model*

Flags
`````

**image**
*Path to image to classify (required)*

tflite
------

*Generate a TFLite file from a frozen graph*

Flags
`````

**output-format**
*TF Lite output format (tflite)

Choices:
tflite
graphviz_dot

*

**quantized**
*Whether or not output arrays are quantized (no)

Choices:
yes
no

*

**quantized-inputs**
*Whether or not input arrays are quantized (no)

Choices:
yes
no

*

train
-----

*Train model from scratch*

Flags
`````

**auto-scale**
*Adjust applicable flags for multi-GPU systems (yes)

Set to 'no' to disable any flag value adjustments.

When this value is 'yes' (the default) the following flags are adjusted on
multi-GPU systems:

- clones

- learning-rate

`clones` is set to the number of available GPUs.

`learning-rate` is adjusted by multiplying its specified value by the
number of GPUs.

Flags are not adjusted on single GPU or CPU only systems.

*

**batch-size**
*Number of examples in each training batch (32)*

**clones**
*Number of model clones (calculated)

This value is automatically set to the number of available GPUs if `auto-
scale` is 'yes'.

When `auto-scale` is 'no' this value can be increased from 1 to train the
model in parallel on multiple GPUs.

*

**learning-rate**
*Initial learning rate (0.001)*

**learning-rate-decay-epochs**
*Number of epochs after which learning rate decays (2.0)*

**learning-rate-decay-factor**
*Learning rate decay factor (0.94)*

**learning-rate-decay-type**
*Method used to decay the learning rate (exponential)

Choices:
exponential
fixed
polynomial

*

**learning-rate-end**
*Minimal learning rate used by polynomial learning rate decay (0.0001)*

**log-save-seconds**
*Frequency of log summary saves in seconds (60)*

**log-steps**
*Frequency of summary logs in steps (100)*

**model-save-seconds**
*Frequency of model saves (checkpoints) in seconds (600)*

**optimizer**
*Optimizer used to train (rmsprop)

Choices:
adadelta
adagrad
adam
ftrl
momentum
rmsprop
sgd

*

**preprocessing**
*Preprocessing to use (default for model)*

**preprocessors**
*Number of preprocessing threads (calculated)

This value is automatically set to logical CPU count / 2 if `auto-scale`
is 'yes'.

When `auto-scale` is 'no' this value can be set to optimize the
preprocessor thread count for the system.

*

**readers**
*Number of parallel data readers (calculated)

This value is automatically set to logical CPU count / 2 if `auto-scale`
is 'yes'.

When `auto-scale` is 'no' this value can be set to optimize data reader
performance for the system.

*

**train-steps**
*Number of steps to train (train indefinitely)*

**weight-decay**
*Decay on the model weights (4e-05)*

transfer-learn
--------------

*Train model using transfer learning*

Flags
`````

**auto-scale**
*Adjust applicable flags for multi-GPU systems (yes)

Set to 'no' to disable any flag value adjustments.

When this value is 'yes' (the default) the following flags are adjusted on
multi-GPU systems:

- clones

- learning-rate

`clones` is set to the number of available GPUs.

`learning-rate` is adjusted by multiplying its specified value by the
number of GPUs.

Flags are not adjusted on single GPU or CPU only systems.

*

**batch-size**
*Number of examples in each training batch (32)*

**clones**
*Number of model clones (calculated)

This value is automatically set to the number of available GPUs if `auto-
scale` is 'yes'.

When `auto-scale` is 'no' this value can be increased from 1 to train the
model in parallel on multiple GPUs.

*

**learning-rate**
*Initial learning rate (0.001)*

**learning-rate-decay-epochs**
*Number of epochs after which learning rate decays (2.0)*

**learning-rate-decay-factor**
*Learning rate decay factor (0.94)*

**learning-rate-decay-type**
*Method used to decay the learning rate (exponential)

Choices:
exponential
fixed
polynomial

*

**learning-rate-end**
*Minimal learning rate used by polynomial learning rate decay (0.0001)*

**log-save-seconds**
*Frequency of log summary saves in seconds (60)*

**log-steps**
*Frequency of summary logs in steps (100)*

**model-save-seconds**
*Frequency of model saves (checkpoints) in seconds (600)*

**optimizer**
*Optimizer used to train (rmsprop)

Choices:
adadelta
adagrad
adam
ftrl
momentum
rmsprop
sgd

*

**preprocessing**
*Preprocessing to use (default for model)*

**preprocessors**
*Number of preprocessing threads (calculated)

This value is automatically set to logical CPU count / 2 if `auto-scale`
is 'yes'.

When `auto-scale` is 'no' this value can be set to optimize the
preprocessor thread count for the system.

*

**readers**
*Number of parallel data readers (calculated)

This value is automatically set to logical CPU count / 2 if `auto-scale`
is 'yes'.

When `auto-scale` is 'no' this value can be set to optimize data reader
performance for the system.

*

**train-steps**
*Number of steps to train (train indefinitely)*

**weight-decay**
*Decay on the model weights (4e-05)*


resnet-v2-152
=============

*TF-Slim ResNet v2 152 layer classifier*

Operations
^^^^^^^^^^

evaluate
--------

*Evaluate a trained model*

Flags
`````

**batch-size**
*Number of examples in each evaluated batch (100)*

**eval-batches**
*Number of batches to evaluate (all available)*

**step**
*Checkpoint step to evaluate (latest checkpoint)*

export-and-freeze
-----------------

*Export an inference graph with checkpoint weights*

Flags
`````

**step**
*Checkpoint step to use for the frozen graph (latest checkpoint)*

finetune
--------

*Finetune a trained model*

Flags
`````

**auto-scale**
*Adjust applicable flags for multi-GPU systems (yes)

Set to 'no' to disable any flag value adjustments.

When this value is 'yes' (the default) the following flags are adjusted on
multi-GPU systems:

- clones

- learning-rate

`clones` is set to the number of available GPUs.

`learning-rate` is adjusted by multiplying its specified value by the
number of GPUs.

Flags are not adjusted on single GPU or CPU only systems.

*

**batch-size**
*Number of examples in each training batch (32)*

**clones**
*Number of model clones (calculated)

This value is automatically set to the number of available GPUs if `auto-
scale` is 'yes'.

When `auto-scale` is 'no' this value can be increased from 1 to train the
model in parallel on multiple GPUs.

*

**learning-rate**
*Initial learning rate (0.0001)*

**learning-rate-decay-epochs**
*Number of epochs after which learning rate decays (2.0)*

**learning-rate-decay-factor**
*Learning rate decay factor (0.94)*

**learning-rate-decay-type**
*Method used to decay the learning rate (exponential)

Choices:
exponential
fixed
polynomial

*

**learning-rate-end**
*Minimal learning rate used by polynomial learning rate decay (0.0001)*

**log-save-seconds**
*Frequency of log summary saves in seconds (60)*

**log-steps**
*Frequency of summary logs in steps (100)*

**model-save-seconds**
*Frequency of model saves (checkpoints) in seconds (600)*

**optimizer**
*Optimizer used to train (rmsprop)

Choices:
adadelta
adagrad
adam
ftrl
momentum
rmsprop
sgd

*

**preprocessing**
*Preprocessing to use (default for model)*

**preprocessors**
*Number of preprocessing threads (calculated)

This value is automatically set to logical CPU count / 2 if `auto-scale`
is 'yes'.

When `auto-scale` is 'no' this value can be set to optimize the
preprocessor thread count for the system.

*

**readers**
*Number of parallel data readers (calculated)

This value is automatically set to logical CPU count / 2 if `auto-scale`
is 'yes'.

When `auto-scale` is 'no' this value can be set to optimize data reader
performance for the system.

*

**train-steps**
*Number of steps to train (train indefinitely)*

**weight-decay**
*Decay on the model weights (4e-05)*

label
-----

*Classify an image using a trained model*

Flags
`````

**image**
*Path to image to classify (required)*

tflite
------

*Generate a TFLite file from a frozen graph*

Flags
`````

**output-format**
*TF Lite output format (tflite)

Choices:
tflite
graphviz_dot

*

**quantized**
*Whether or not output arrays are quantized (no)

Choices:
yes
no

*

**quantized-inputs**
*Whether or not input arrays are quantized (no)

Choices:
yes
no

*

train
-----

*Train model from scratch*

Flags
`````

**auto-scale**
*Adjust applicable flags for multi-GPU systems (yes)

Set to 'no' to disable any flag value adjustments.

When this value is 'yes' (the default) the following flags are adjusted on
multi-GPU systems:

- clones

- learning-rate

`clones` is set to the number of available GPUs.

`learning-rate` is adjusted by multiplying its specified value by the
number of GPUs.

Flags are not adjusted on single GPU or CPU only systems.

*

**batch-size**
*Number of examples in each training batch (32)*

**clones**
*Number of model clones (calculated)

This value is automatically set to the number of available GPUs if `auto-
scale` is 'yes'.

When `auto-scale` is 'no' this value can be increased from 1 to train the
model in parallel on multiple GPUs.

*

**learning-rate**
*Initial learning rate (0.001)*

**learning-rate-decay-epochs**
*Number of epochs after which learning rate decays (2.0)*

**learning-rate-decay-factor**
*Learning rate decay factor (0.94)*

**learning-rate-decay-type**
*Method used to decay the learning rate (exponential)

Choices:
exponential
fixed
polynomial

*

**learning-rate-end**
*Minimal learning rate used by polynomial learning rate decay (0.0001)*

**log-save-seconds**
*Frequency of log summary saves in seconds (60)*

**log-steps**
*Frequency of summary logs in steps (100)*

**model-save-seconds**
*Frequency of model saves (checkpoints) in seconds (600)*

**optimizer**
*Optimizer used to train (rmsprop)

Choices:
adadelta
adagrad
adam
ftrl
momentum
rmsprop
sgd

*

**preprocessing**
*Preprocessing to use (default for model)*

**preprocessors**
*Number of preprocessing threads (calculated)

This value is automatically set to logical CPU count / 2 if `auto-scale`
is 'yes'.

When `auto-scale` is 'no' this value can be set to optimize the
preprocessor thread count for the system.

*

**readers**
*Number of parallel data readers (calculated)

This value is automatically set to logical CPU count / 2 if `auto-scale`
is 'yes'.

When `auto-scale` is 'no' this value can be set to optimize data reader
performance for the system.

*

**train-steps**
*Number of steps to train (train indefinitely)*

**weight-decay**
*Decay on the model weights (4e-05)*

transfer-learn
--------------

*Train model using transfer learning*

Flags
`````

**auto-scale**
*Adjust applicable flags for multi-GPU systems (yes)

Set to 'no' to disable any flag value adjustments.

When this value is 'yes' (the default) the following flags are adjusted on
multi-GPU systems:

- clones

- learning-rate

`clones` is set to the number of available GPUs.

`learning-rate` is adjusted by multiplying its specified value by the
number of GPUs.

Flags are not adjusted on single GPU or CPU only systems.

*

**batch-size**
*Number of examples in each training batch (32)*

**clones**
*Number of model clones (calculated)

This value is automatically set to the number of available GPUs if `auto-
scale` is 'yes'.

When `auto-scale` is 'no' this value can be increased from 1 to train the
model in parallel on multiple GPUs.

*

**learning-rate**
*Initial learning rate (0.001)*

**learning-rate-decay-epochs**
*Number of epochs after which learning rate decays (2.0)*

**learning-rate-decay-factor**
*Learning rate decay factor (0.94)*

**learning-rate-decay-type**
*Method used to decay the learning rate (exponential)

Choices:
exponential
fixed
polynomial

*

**learning-rate-end**
*Minimal learning rate used by polynomial learning rate decay (0.0001)*

**log-save-seconds**
*Frequency of log summary saves in seconds (60)*

**log-steps**
*Frequency of summary logs in steps (100)*

**model-save-seconds**
*Frequency of model saves (checkpoints) in seconds (600)*

**optimizer**
*Optimizer used to train (rmsprop)

Choices:
adadelta
adagrad
adam
ftrl
momentum
rmsprop
sgd

*

**preprocessing**
*Preprocessing to use (default for model)*

**preprocessors**
*Number of preprocessing threads (calculated)

This value is automatically set to logical CPU count / 2 if `auto-scale`
is 'yes'.

When `auto-scale` is 'no' this value can be set to optimize the
preprocessor thread count for the system.

*

**readers**
*Number of parallel data readers (calculated)

This value is automatically set to logical CPU count / 2 if `auto-scale`
is 'yes'.

When `auto-scale` is 'no' this value can be set to optimize data reader
performance for the system.

*

**train-steps**
*Number of steps to train (train indefinitely)*

**weight-decay**
*Decay on the model weights (4e-05)*


resnet-v2-50
============

*TF-Slim ResNet v2 50 layer classifier*

Operations
^^^^^^^^^^

evaluate
--------

*Evaluate a trained model*

Flags
`````

**batch-size**
*Number of examples in each evaluated batch (100)*

**eval-batches**
*Number of batches to evaluate (all available)*

**step**
*Checkpoint step to evaluate (latest checkpoint)*

export-and-freeze
-----------------

*Export an inference graph with checkpoint weights*

Flags
`````

**step**
*Checkpoint step to use for the frozen graph (latest checkpoint)*

finetune
--------

*Finetune a trained model*

Flags
`````

**auto-scale**
*Adjust applicable flags for multi-GPU systems (yes)

Set to 'no' to disable any flag value adjustments.

When this value is 'yes' (the default) the following flags are adjusted on
multi-GPU systems:

- clones

- learning-rate

`clones` is set to the number of available GPUs.

`learning-rate` is adjusted by multiplying its specified value by the
number of GPUs.

Flags are not adjusted on single GPU or CPU only systems.

*

**batch-size**
*Number of examples in each training batch (32)*

**clones**
*Number of model clones (calculated)

This value is automatically set to the number of available GPUs if `auto-
scale` is 'yes'.

When `auto-scale` is 'no' this value can be increased from 1 to train the
model in parallel on multiple GPUs.

*

**learning-rate**
*Initial learning rate (0.0001)*

**learning-rate-decay-epochs**
*Number of epochs after which learning rate decays (2.0)*

**learning-rate-decay-factor**
*Learning rate decay factor (0.94)*

**learning-rate-decay-type**
*Method used to decay the learning rate (exponential)

Choices:
exponential
fixed
polynomial

*

**learning-rate-end**
*Minimal learning rate used by polynomial learning rate decay (0.0001)*

**log-save-seconds**
*Frequency of log summary saves in seconds (60)*

**log-steps**
*Frequency of summary logs in steps (100)*

**model-save-seconds**
*Frequency of model saves (checkpoints) in seconds (600)*

**optimizer**
*Optimizer used to train (rmsprop)

Choices:
adadelta
adagrad
adam
ftrl
momentum
rmsprop
sgd

*

**preprocessing**
*Preprocessing to use (default for model)*

**preprocessors**
*Number of preprocessing threads (calculated)

This value is automatically set to logical CPU count / 2 if `auto-scale`
is 'yes'.

When `auto-scale` is 'no' this value can be set to optimize the
preprocessor thread count for the system.

*

**readers**
*Number of parallel data readers (calculated)

This value is automatically set to logical CPU count / 2 if `auto-scale`
is 'yes'.

When `auto-scale` is 'no' this value can be set to optimize data reader
performance for the system.

*

**train-steps**
*Number of steps to train (train indefinitely)*

**weight-decay**
*Decay on the model weights (4e-05)*

label
-----

*Classify an image using a trained model*

Flags
`````

**image**
*Path to image to classify (required)*

tflite
------

*Generate a TFLite file from a frozen graph*

Flags
`````

**output-format**
*TF Lite output format (tflite)

Choices:
tflite
graphviz_dot

*

**quantized**
*Whether or not output arrays are quantized (no)

Choices:
yes
no

*

**quantized-inputs**
*Whether or not input arrays are quantized (no)

Choices:
yes
no

*

train
-----

*Train model from scratch*

Flags
`````

**auto-scale**
*Adjust applicable flags for multi-GPU systems (yes)

Set to 'no' to disable any flag value adjustments.

When this value is 'yes' (the default) the following flags are adjusted on
multi-GPU systems:

- clones

- learning-rate

`clones` is set to the number of available GPUs.

`learning-rate` is adjusted by multiplying its specified value by the
number of GPUs.

Flags are not adjusted on single GPU or CPU only systems.

*

**batch-size**
*Number of examples in each training batch (32)*

**clones**
*Number of model clones (calculated)

This value is automatically set to the number of available GPUs if `auto-
scale` is 'yes'.

When `auto-scale` is 'no' this value can be increased from 1 to train the
model in parallel on multiple GPUs.

*

**learning-rate**
*Initial learning rate (0.001)*

**learning-rate-decay-epochs**
*Number of epochs after which learning rate decays (2.0)*

**learning-rate-decay-factor**
*Learning rate decay factor (0.94)*

**learning-rate-decay-type**
*Method used to decay the learning rate (exponential)

Choices:
exponential
fixed
polynomial

*

**learning-rate-end**
*Minimal learning rate used by polynomial learning rate decay (0.0001)*

**log-save-seconds**
*Frequency of log summary saves in seconds (60)*

**log-steps**
*Frequency of summary logs in steps (100)*

**model-save-seconds**
*Frequency of model saves (checkpoints) in seconds (600)*

**optimizer**
*Optimizer used to train (rmsprop)

Choices:
adadelta
adagrad
adam
ftrl
momentum
rmsprop
sgd

*

**preprocessing**
*Preprocessing to use (default for model)*

**preprocessors**
*Number of preprocessing threads (calculated)

This value is automatically set to logical CPU count / 2 if `auto-scale`
is 'yes'.

When `auto-scale` is 'no' this value can be set to optimize the
preprocessor thread count for the system.

*

**readers**
*Number of parallel data readers (calculated)

This value is automatically set to logical CPU count / 2 if `auto-scale`
is 'yes'.

When `auto-scale` is 'no' this value can be set to optimize data reader
performance for the system.

*

**train-steps**
*Number of steps to train (train indefinitely)*

**weight-decay**
*Decay on the model weights (4e-05)*

transfer-learn
--------------

*Train model using transfer learning*

Flags
`````

**auto-scale**
*Adjust applicable flags for multi-GPU systems (yes)

Set to 'no' to disable any flag value adjustments.

When this value is 'yes' (the default) the following flags are adjusted on
multi-GPU systems:

- clones

- learning-rate

`clones` is set to the number of available GPUs.

`learning-rate` is adjusted by multiplying its specified value by the
number of GPUs.

Flags are not adjusted on single GPU or CPU only systems.

*

**batch-size**
*Number of examples in each training batch (32)*

**clones**
*Number of model clones (calculated)

This value is automatically set to the number of available GPUs if `auto-
scale` is 'yes'.

When `auto-scale` is 'no' this value can be increased from 1 to train the
model in parallel on multiple GPUs.

*

**learning-rate**
*Initial learning rate (0.001)*

**learning-rate-decay-epochs**
*Number of epochs after which learning rate decays (2.0)*

**learning-rate-decay-factor**
*Learning rate decay factor (0.94)*

**learning-rate-decay-type**
*Method used to decay the learning rate (exponential)

Choices:
exponential
fixed
polynomial

*

**learning-rate-end**
*Minimal learning rate used by polynomial learning rate decay (0.0001)*

**log-save-seconds**
*Frequency of log summary saves in seconds (60)*

**log-steps**
*Frequency of summary logs in steps (100)*

**model-save-seconds**
*Frequency of model saves (checkpoints) in seconds (600)*

**optimizer**
*Optimizer used to train (rmsprop)

Choices:
adadelta
adagrad
adam
ftrl
momentum
rmsprop
sgd

*

**preprocessing**
*Preprocessing to use (default for model)*

**preprocessors**
*Number of preprocessing threads (calculated)

This value is automatically set to logical CPU count / 2 if `auto-scale`
is 'yes'.

When `auto-scale` is 'no' this value can be set to optimize the
preprocessor thread count for the system.

*

**readers**
*Number of parallel data readers (calculated)

This value is automatically set to logical CPU count / 2 if `auto-scale`
is 'yes'.

When `auto-scale` is 'no' this value can be set to optimize data reader
performance for the system.

*

**train-steps**
*Number of steps to train (train indefinitely)*

**weight-decay**
*Decay on the model weights (4e-05)*


vgg-16
======

*TF-Slim VGG 16 classifier*

Operations
^^^^^^^^^^

evaluate
--------

*Evaluate a trained model*

Flags
`````

**batch-size**
*Number of examples in each evaluated batch (100)*

**eval-batches**
*Number of batches to evaluate (all available)*

**step**
*Checkpoint step to evaluate (latest checkpoint)*

export-and-freeze
-----------------

*Export an inference graph with checkpoint weights*

Flags
`````

**step**
*Checkpoint step to use for the frozen graph (latest checkpoint)*

finetune
--------

*Finetune a trained model*

Flags
`````

**auto-scale**
*Adjust applicable flags for multi-GPU systems (yes)

Set to 'no' to disable any flag value adjustments.

When this value is 'yes' (the default) the following flags are adjusted on
multi-GPU systems:

- clones

- learning-rate

`clones` is set to the number of available GPUs.

`learning-rate` is adjusted by multiplying its specified value by the
number of GPUs.

Flags are not adjusted on single GPU or CPU only systems.

*

**batch-size**
*Number of examples in each training batch (32)*

**clones**
*Number of model clones (calculated)

This value is automatically set to the number of available GPUs if `auto-
scale` is 'yes'.

When `auto-scale` is 'no' this value can be increased from 1 to train the
model in parallel on multiple GPUs.

*

**learning-rate**
*Initial learning rate (0.0001)*

**learning-rate-decay-epochs**
*Number of epochs after which learning rate decays (2.0)*

**learning-rate-decay-factor**
*Learning rate decay factor (0.94)*

**learning-rate-decay-type**
*Method used to decay the learning rate (exponential)

Choices:
exponential
fixed
polynomial

*

**learning-rate-end**
*Minimal learning rate used by polynomial learning rate decay (0.0001)*

**log-save-seconds**
*Frequency of log summary saves in seconds (60)*

**log-steps**
*Frequency of summary logs in steps (100)*

**model-save-seconds**
*Frequency of model saves (checkpoints) in seconds (600)*

**optimizer**
*Optimizer used to train (rmsprop)

Choices:
adadelta
adagrad
adam
ftrl
momentum
rmsprop
sgd

*

**preprocessing**
*Preprocessing to use (default for model)*

**preprocessors**
*Number of preprocessing threads (calculated)

This value is automatically set to logical CPU count / 2 if `auto-scale`
is 'yes'.

When `auto-scale` is 'no' this value can be set to optimize the
preprocessor thread count for the system.

*

**readers**
*Number of parallel data readers (calculated)

This value is automatically set to logical CPU count / 2 if `auto-scale`
is 'yes'.

When `auto-scale` is 'no' this value can be set to optimize data reader
performance for the system.

*

**train-steps**
*Number of steps to train (train indefinitely)*

**weight-decay**
*Decay on the model weights (4e-05)*

label
-----

*Classify an image using a trained model*

Flags
`````

**image**
*Path to image to classify (required)*

tflite
------

*Generate a TFLite file from a frozen graph*

Flags
`````

**output-format**
*TF Lite output format (tflite)

Choices:
tflite
graphviz_dot

*

**quantized**
*Whether or not output arrays are quantized (no)

Choices:
yes
no

*

**quantized-inputs**
*Whether or not input arrays are quantized (no)

Choices:
yes
no

*

train
-----

*Train model from scratch*

Flags
`````

**auto-scale**
*Adjust applicable flags for multi-GPU systems (yes)

Set to 'no' to disable any flag value adjustments.

When this value is 'yes' (the default) the following flags are adjusted on
multi-GPU systems:

- clones

- learning-rate

`clones` is set to the number of available GPUs.

`learning-rate` is adjusted by multiplying its specified value by the
number of GPUs.

Flags are not adjusted on single GPU or CPU only systems.

*

**batch-size**
*Number of examples in each training batch (32)*

**clones**
*Number of model clones (calculated)

This value is automatically set to the number of available GPUs if `auto-
scale` is 'yes'.

When `auto-scale` is 'no' this value can be increased from 1 to train the
model in parallel on multiple GPUs.

*

**learning-rate**
*Initial learning rate (0.001)*

**learning-rate-decay-epochs**
*Number of epochs after which learning rate decays (2.0)*

**learning-rate-decay-factor**
*Learning rate decay factor (0.94)*

**learning-rate-decay-type**
*Method used to decay the learning rate (exponential)

Choices:
exponential
fixed
polynomial

*

**learning-rate-end**
*Minimal learning rate used by polynomial learning rate decay (0.0001)*

**log-save-seconds**
*Frequency of log summary saves in seconds (60)*

**log-steps**
*Frequency of summary logs in steps (100)*

**model-save-seconds**
*Frequency of model saves (checkpoints) in seconds (600)*

**optimizer**
*Optimizer used to train (rmsprop)

Choices:
adadelta
adagrad
adam
ftrl
momentum
rmsprop
sgd

*

**preprocessing**
*Preprocessing to use (default for model)*

**preprocessors**
*Number of preprocessing threads (calculated)

This value is automatically set to logical CPU count / 2 if `auto-scale`
is 'yes'.

When `auto-scale` is 'no' this value can be set to optimize the
preprocessor thread count for the system.

*

**readers**
*Number of parallel data readers (calculated)

This value is automatically set to logical CPU count / 2 if `auto-scale`
is 'yes'.

When `auto-scale` is 'no' this value can be set to optimize data reader
performance for the system.

*

**train-steps**
*Number of steps to train (train indefinitely)*

**weight-decay**
*Decay on the model weights (4e-05)*

transfer-learn
--------------

*Train model using transfer learning*

Flags
`````

**auto-scale**
*Adjust applicable flags for multi-GPU systems (yes)

Set to 'no' to disable any flag value adjustments.

When this value is 'yes' (the default) the following flags are adjusted on
multi-GPU systems:

- clones

- learning-rate

`clones` is set to the number of available GPUs.

`learning-rate` is adjusted by multiplying its specified value by the
number of GPUs.

Flags are not adjusted on single GPU or CPU only systems.

*

**batch-size**
*Number of examples in each training batch (32)*

**clones**
*Number of model clones (calculated)

This value is automatically set to the number of available GPUs if `auto-
scale` is 'yes'.

When `auto-scale` is 'no' this value can be increased from 1 to train the
model in parallel on multiple GPUs.

*

**learning-rate**
*Initial learning rate (0.001)*

**learning-rate-decay-epochs**
*Number of epochs after which learning rate decays (2.0)*

**learning-rate-decay-factor**
*Learning rate decay factor (0.94)*

**learning-rate-decay-type**
*Method used to decay the learning rate (exponential)

Choices:
exponential
fixed
polynomial

*

**learning-rate-end**
*Minimal learning rate used by polynomial learning rate decay (0.0001)*

**log-save-seconds**
*Frequency of log summary saves in seconds (60)*

**log-steps**
*Frequency of summary logs in steps (100)*

**model-save-seconds**
*Frequency of model saves (checkpoints) in seconds (600)*

**optimizer**
*Optimizer used to train (rmsprop)

Choices:
adadelta
adagrad
adam
ftrl
momentum
rmsprop
sgd

*

**preprocessing**
*Preprocessing to use (default for model)*

**preprocessors**
*Number of preprocessing threads (calculated)

This value is automatically set to logical CPU count / 2 if `auto-scale`
is 'yes'.

When `auto-scale` is 'no' this value can be set to optimize the
preprocessor thread count for the system.

*

**readers**
*Number of parallel data readers (calculated)

This value is automatically set to logical CPU count / 2 if `auto-scale`
is 'yes'.

When `auto-scale` is 'no' this value can be set to optimize data reader
performance for the system.

*

**train-steps**
*Number of steps to train (train indefinitely)*

**weight-decay**
*Decay on the model weights (4e-05)*


vgg-19
======

*TF-Slim VGG 19 classifier*

Operations
^^^^^^^^^^

evaluate
--------

*Evaluate a trained model*

Flags
`````

**batch-size**
*Number of examples in each evaluated batch (100)*

**eval-batches**
*Number of batches to evaluate (all available)*

**step**
*Checkpoint step to evaluate (latest checkpoint)*

export-and-freeze
-----------------

*Export an inference graph with checkpoint weights*

Flags
`````

**step**
*Checkpoint step to use for the frozen graph (latest checkpoint)*

finetune
--------

*Finetune a trained model*

Flags
`````

**auto-scale**
*Adjust applicable flags for multi-GPU systems (yes)

Set to 'no' to disable any flag value adjustments.

When this value is 'yes' (the default) the following flags are adjusted on
multi-GPU systems:

- clones

- learning-rate

`clones` is set to the number of available GPUs.

`learning-rate` is adjusted by multiplying its specified value by the
number of GPUs.

Flags are not adjusted on single GPU or CPU only systems.

*

**batch-size**
*Number of examples in each training batch (32)*

**clones**
*Number of model clones (calculated)

This value is automatically set to the number of available GPUs if `auto-
scale` is 'yes'.

When `auto-scale` is 'no' this value can be increased from 1 to train the
model in parallel on multiple GPUs.

*

**learning-rate**
*Initial learning rate (0.0001)*

**learning-rate-decay-epochs**
*Number of epochs after which learning rate decays (2.0)*

**learning-rate-decay-factor**
*Learning rate decay factor (0.94)*

**learning-rate-decay-type**
*Method used to decay the learning rate (exponential)

Choices:
exponential
fixed
polynomial

*

**learning-rate-end**
*Minimal learning rate used by polynomial learning rate decay (0.0001)*

**log-save-seconds**
*Frequency of log summary saves in seconds (60)*

**log-steps**
*Frequency of summary logs in steps (100)*

**model-save-seconds**
*Frequency of model saves (checkpoints) in seconds (600)*

**optimizer**
*Optimizer used to train (rmsprop)

Choices:
adadelta
adagrad
adam
ftrl
momentum
rmsprop
sgd

*

**preprocessing**
*Preprocessing to use (default for model)*

**preprocessors**
*Number of preprocessing threads (calculated)

This value is automatically set to logical CPU count / 2 if `auto-scale`
is 'yes'.

When `auto-scale` is 'no' this value can be set to optimize the
preprocessor thread count for the system.

*

**readers**
*Number of parallel data readers (calculated)

This value is automatically set to logical CPU count / 2 if `auto-scale`
is 'yes'.

When `auto-scale` is 'no' this value can be set to optimize data reader
performance for the system.

*

**train-steps**
*Number of steps to train (train indefinitely)*

**weight-decay**
*Decay on the model weights (4e-05)*

label
-----

*Classify an image using a trained model*

Flags
`````

**image**
*Path to image to classify (required)*

tflite
------

*Generate a TFLite file from a frozen graph*

Flags
`````

**output-format**
*TF Lite output format (tflite)

Choices:
tflite
graphviz_dot

*

**quantized**
*Whether or not output arrays are quantized (no)

Choices:
yes
no

*

**quantized-inputs**
*Whether or not input arrays are quantized (no)

Choices:
yes
no

*

train
-----

*Train model from scratch*

Flags
`````

**auto-scale**
*Adjust applicable flags for multi-GPU systems (yes)

Set to 'no' to disable any flag value adjustments.

When this value is 'yes' (the default) the following flags are adjusted on
multi-GPU systems:

- clones

- learning-rate

`clones` is set to the number of available GPUs.

`learning-rate` is adjusted by multiplying its specified value by the
number of GPUs.

Flags are not adjusted on single GPU or CPU only systems.

*

**batch-size**
*Number of examples in each training batch (32)*

**clones**
*Number of model clones (calculated)

This value is automatically set to the number of available GPUs if `auto-
scale` is 'yes'.

When `auto-scale` is 'no' this value can be increased from 1 to train the
model in parallel on multiple GPUs.

*

**learning-rate**
*Initial learning rate (0.001)*

**learning-rate-decay-epochs**
*Number of epochs after which learning rate decays (2.0)*

**learning-rate-decay-factor**
*Learning rate decay factor (0.94)*

**learning-rate-decay-type**
*Method used to decay the learning rate (exponential)

Choices:
exponential
fixed
polynomial

*

**learning-rate-end**
*Minimal learning rate used by polynomial learning rate decay (0.0001)*

**log-save-seconds**
*Frequency of log summary saves in seconds (60)*

**log-steps**
*Frequency of summary logs in steps (100)*

**model-save-seconds**
*Frequency of model saves (checkpoints) in seconds (600)*

**optimizer**
*Optimizer used to train (rmsprop)

Choices:
adadelta
adagrad
adam
ftrl
momentum
rmsprop
sgd

*

**preprocessing**
*Preprocessing to use (default for model)*

**preprocessors**
*Number of preprocessing threads (calculated)

This value is automatically set to logical CPU count / 2 if `auto-scale`
is 'yes'.

When `auto-scale` is 'no' this value can be set to optimize the
preprocessor thread count for the system.

*

**readers**
*Number of parallel data readers (calculated)

This value is automatically set to logical CPU count / 2 if `auto-scale`
is 'yes'.

When `auto-scale` is 'no' this value can be set to optimize data reader
performance for the system.

*

**train-steps**
*Number of steps to train (train indefinitely)*

**weight-decay**
*Decay on the model weights (4e-05)*

transfer-learn
--------------

*Train model using transfer learning*

Flags
`````

**auto-scale**
*Adjust applicable flags for multi-GPU systems (yes)

Set to 'no' to disable any flag value adjustments.

When this value is 'yes' (the default) the following flags are adjusted on
multi-GPU systems:

- clones

- learning-rate

`clones` is set to the number of available GPUs.

`learning-rate` is adjusted by multiplying its specified value by the
number of GPUs.

Flags are not adjusted on single GPU or CPU only systems.

*

**batch-size**
*Number of examples in each training batch (32)*

**clones**
*Number of model clones (calculated)

This value is automatically set to the number of available GPUs if `auto-
scale` is 'yes'.

When `auto-scale` is 'no' this value can be increased from 1 to train the
model in parallel on multiple GPUs.

*

**learning-rate**
*Initial learning rate (0.001)*

**learning-rate-decay-epochs**
*Number of epochs after which learning rate decays (2.0)*

**learning-rate-decay-factor**
*Learning rate decay factor (0.94)*

**learning-rate-decay-type**
*Method used to decay the learning rate (exponential)

Choices:
exponential
fixed
polynomial

*

**learning-rate-end**
*Minimal learning rate used by polynomial learning rate decay (0.0001)*

**log-save-seconds**
*Frequency of log summary saves in seconds (60)*

**log-steps**
*Frequency of summary logs in steps (100)*

**model-save-seconds**
*Frequency of model saves (checkpoints) in seconds (600)*

**optimizer**
*Optimizer used to train (rmsprop)

Choices:
adadelta
adagrad
adam
ftrl
momentum
rmsprop
sgd

*

**preprocessing**
*Preprocessing to use (default for model)*

**preprocessors**
*Number of preprocessing threads (calculated)

This value is automatically set to logical CPU count / 2 if `auto-scale`
is 'yes'.

When `auto-scale` is 'no' this value can be set to optimize the
preprocessor thread count for the system.

*

**readers**
*Number of parallel data readers (calculated)

This value is automatically set to logical CPU count / 2 if `auto-scale`
is 'yes'.

When `auto-scale` is 'no' this value can be set to optimize data reader
performance for the system.

*

**train-steps**
*Number of steps to train (train indefinitely)*

**weight-decay**
*Decay on the model weights (4e-05)*